feat: better tui menu option gating for models that support custom voices (qwen) and models that do not support instructions

11 files changed, 535 insertions, 171 deletions

diff --git a/README.md b/README.md
index 3a6419b..9e9a6b2 100644
--- a/README.md
+++ b/README.md
@@ -68,7 +68,8 @@ Everything the TUI does can also be scripted with flags: `python audiobook.py --
 | `--model <id>`                     | `audiocpp`: Choose the model from `server.json`                                                                                                                                                                                                               |
 | `--instructions "..."`             | `audiocpp`: voice design or style instruction. Required for voice design models (`vdes`)                                                                           |
 | `--option KEY=VALUE`               | `audiocpp`: Some models support custom options (e.g. `emotion=netural`) that can be passed with this flag                                                                                                                                                          |
-| `--voice <name>`                   | `audiocpp`, `faster`: Server-side voice to request                                                                                                                                                                                                                 |
+| `--voice <name>`                   | `audiocpp`, `faster`: Server-side voice to request (`audiocpp`: a `voice_preset`/`voice_dir` entry for cloning; with `AUDIOCPP_CLONE_MODEL_ID` set it reroutes to the clone model — typically the Qwen Base model).                                                                  |
+| `--speaker <name>`                 | `audiocpp`: Built-in Qwen3-TTS CustomVoice speaker (e.g. `Vivian`, `Ryan`, `Uncle_Fu`). Selects speaker mode on the CustomVoice model entry; mutually exclusive with `--voice`.                                                                                              |
 | `--clone <path>`                   | `qwen`: Reference audio (`wav`) for voice cloning.                                                                                                                                                                                                                 |
 | `--transcription "..."`            | `qwen`: Override whisper auto-transcription with manual audio transcript.                                                                                                                                                                                          |
 | `--no-transcription`               | `qwen`: Skip auto-transcription of the reference audio.                                                                                                                                                                                                            |
diff --git a/app/backends/qwen.py b/app/backends/qwen.py
index 7fba7e0..be7ebfe 100644
--- a/app/backends/qwen.py
+++ b/app/backends/qwen.py
@@ -30,6 +30,7 @@ from backends import (
     servers,
 )
 from converter import config
+from converter.tts import QWEN3_TTS_SPEAKERS
 from ui import taskview, tui
 
 QWEN_PIP_PKG = "qwen-tts"
@@ -38,9 +39,10 @@ QWEN_BASE_MODEL = "Qwen/Qwen3-TTS-12Hz-1.7B-Base"
 DEFAULT_CUSTOM_PORT = 7860
 DEFAULT_CLONE_PORT = 7861
 
-# Built-in CustomVoice speakers (see app/converter/config.py SPEAKER).
-QWEN_SPEAKERS = ("Vivian", "Serena", "Uncle_Fu", "Dylan", "Eric", "Ryan",
-                 "Aiden", "Ono_Anna", "Sohee")
+# Built-in CustomVoice speakers (see app/converter/config.py SPEAKER). The
+# canonical list lives in converter.tts (shared with the audiocpp backend's
+# Convert-form Speaker picker).
+QWEN_SPEAKERS = QWEN3_TTS_SPEAKERS
 
 
 def _is_installed() -> bool:
diff --git a/app/converter/converter.py b/app/converter/converter.py
index b2c2923..6e677c8 100644
--- a/app/converter/converter.py
+++ b/app/converter/converter.py
@@ -185,6 +185,7 @@ class AudiobookConverter:
                  instructions: Optional[str] = None,
                  request_options: Optional[Dict[str, str]] = None,
                  api_url: Optional[str] = None,
+                 speaker: Optional[str] = None,
                  progress: Optional[Callable[[dict], None]] = None,
                  cancel=None):
         if speed <= 0:
@@ -205,6 +206,7 @@ class AudiobookConverter:
         self.output_format = output_format
         self.backend = backend
         self.voice = voice
+        self.speaker = speaker
         self.debug = bool(debug)
         # Voice design / style instruction and free-form request options
         # (audio.cpp only): forwarded to AudioCppTTSClient, which validates
@@ -217,16 +219,18 @@ class AudiobookConverter:
             # configured on the server, so no local reference audio is needed.
             self.tts = FasterTTSClient(voice=voice, api_url=api_url)
         elif backend == BACKEND_AUDIOCPP:
-            # Speaker mode (no voice) uses a built-in CustomVoice speaker;
-            # an explicit voice selects a server-side preset (cloning).
-            # model_id overrides AUDIOCPP_MODEL_ID for multi-model servers;
-            # instructions describe or style the voice, request_options pass
-            # per-model controls through to the server.
+            # Speaker mode (--speaker or no flag on a CustomVoice entry) uses
+            # a built-in speaker name; an explicit --voice selects a
+            # server-side preset (cloning). model_id overrides
+            # AUDIOCPP_MODEL_ID for multi-model servers; instructions describe
+            # or style the voice, request_options pass per-model controls
+            # through to the server.
             self.tts = AudioCppTTSClient(voice=voice, language=self.language,
                                          model_id=model_id,
                                          instructions=instructions,
                                          request_options=self.request_options,
-                                         api_url=api_url)
+                                         api_url=api_url,
+                                         speaker=speaker)
         else:
             self.tts = QwenTTSClient(
                 voice_mode=voice_mode,
@@ -290,22 +294,24 @@ class AudiobookConverter:
         """Narrator name used in output file names (see compute_narrator_tag)."""
         return self.compute_narrator_tag(
             self.backend, self.voice, self.voice_mode,
-            self.voice_clone_ref_audio, self.instructions)
+            self.voice_clone_ref_audio, self.instructions, self.speaker)
 
     @staticmethod
     def compute_narrator_tag(backend: str, voice: Optional[str],
                              voice_mode: str,
                              voice_clone_ref_audio: Optional[str],
-                             instructions: Optional[str] = None) -> str:
+                             instructions: Optional[str] = None,
+                             speaker: Optional[str] = None) -> str:
         """Narrator name used in output file names, without a server connection.
 
         Custom voice mode uses the built-in speaker's display name; voice
         clone mode uses the reference audio file's stem; the faster and
-        audiocpp backends use the server-side voice name (falling back to
-        the built-in speaker for the audiocpp backend's speaker mode). An
-        instruction without a voice (voice design, or instruction-defined
-        voices on families without built-in speakers) uses "designed".
-        Spaces become underscores (e.g. "Uncle Fu" -> "Uncle_Fu").
+        audiocpp backends use the server-side voice name (or, for audiocpp's
+        speaker mode, the selected built-in CustomVoice speaker, falling back
+        to the configured one). An instruction without a voice (voice design,
+        or instruction-defined voices on families without built-in speakers)
+        uses "designed". Spaces become underscores (e.g. "Uncle Fu" ->
+        "Uncle_Fu").
 
         Pure (no I/O, no server) so the pre-flight overwrite check can
         compute the exact output names a run would produce before spending
@@ -316,6 +322,8 @@ class AudiobookConverter:
         elif backend == BACKEND_AUDIOCPP:
             if voice:
                 narrator = voice
+            elif speaker:
+                narrator = speaker
             elif instructions:
                 # The voice comes from the instruction, not a speaker name.
                 narrator = "designed"
@@ -672,6 +680,9 @@ class AudiobookConverter:
             if self.voice:
                 self._say("Backend: audio.cpp (voice cloning, reference configured on server)")
                 self._say(f"Voice: {self.voice}")
+            elif self.speaker:
+                self._say("Backend: audio.cpp (custom voice, built-in speaker)")
+                self._say(f"Speaker: {self.speaker}")
             elif self.instructions:
                 self._say("Backend: audio.cpp (voice from --instructions description)")
                 self._say(f"Instruction: {self.instructions}")
@@ -715,6 +726,7 @@ class AudiobookConverter:
                              voice_clone_ref_audio: Optional[str],
                              output_format: str,
                              instructions: Optional[str] = None,
+                             speaker: Optional[str] = None,
                              confirm: Optional[Callable[[str, bool], bool]] = None,
                              ) -> Tuple[List[Path], List[Tuple[Path, str]]]:
         """Discover books and ask every overwrite question up front.
@@ -746,7 +758,8 @@ class AudiobookConverter:
         # starts, so the rest of the run is unattended.
         planned: List[Tuple[Path, str]] = []
         narrator_tag = AudiobookConverter.compute_narrator_tag(
-            backend, voice, voice_mode, voice_clone_ref_audio, instructions)
+            backend, voice, voice_mode, voice_clone_ref_audio, instructions,
+            speaker)
         for book_file in book_files:
             output_name = book_file.stem
             if stem_counts[book_file.stem] > 1:
@@ -782,7 +795,7 @@ class AudiobookConverter:
             book_files, planned = AudiobookConverter.preflight_overwrites(
                 self.backend, self.voice, self.voice_mode,
                 self.voice_clone_ref_audio, self.output_format,
-                self.instructions)
+                self.instructions, self.speaker)
 
         if not book_files:
             self._say(f"[INFO] No supported files found in {BOOKS_FOLDER}")
diff --git a/app/converter/tts.py b/app/converter/tts.py
index 0667ec3..a88697e 100644
--- a/app/converter/tts.py
+++ b/app/converter/tts.py
@@ -121,32 +121,51 @@ AUDIOCPP_LANG_DISPLAY = "display"  # Qwen display names, e.g. "English"
 AUDIOCPP_LANG_ISO = "iso"          # ISO 639-1 codes, e.g. "en"
 AUDIOCPP_LANG_OMIT = "omit"        # no language field; the model detects it
 
-# The only family with a built-in speaker mode (CustomVoice speaker names
-# plus the INSTRUCT style prompt). Every other family is clone-only: the
-# voice comes from a server-side preset requested with --voice.
+# The Qwen3-TTS family. Unlike every other family (one model type each),
+# qwen3_tts hosts several model *types* under one family id, distinguished
+# only by the server entry's id/task: the CustomVoice model (built-in
+# speakers, e.g. Vivian/Ryan), the Base model (voice cloning via a
+# server-side preset), and the VoiceDesign model (task "vdes"). The
+# per-entry voice capability below (audiocpp_entry_voice_capability)
+# resolves which is which, driving both the Convert form (which voice
+# list to show) and the converter's mode selection.
 AUDIOCPP_FAMILY_QWEN3_TTS = "qwen3_tts"
 
 # Server model entry tasks this client can synthesize audiobooks with,
-# taken from GET /v1/models (the "task" field of each entry; servers that
-# predate the field reported TTS models only, so a missing task is treated
-# as "tts"). "vdes" entries are voice design models: the voice is described
-# with --instructions instead of coming from a speaker or a reference clip.
-# Entries with any other task (asr, vc, diar, ...) are rejected at connect
-# time with a hint to pick a synthesis entry.
+# taken from GET /v1/models (the "task" field of each entry; a missing task
+# is treated as "tts" — a harmless generic default). "vdes" entries are
+# voice design models: the voice is described with --instructions instead
+# of coming from a speaker or a reference clip. Entries with any other task
+# (asr, vc, diar, ...) are rejected at connect time with a hint to pick a
+# synthesis entry.
 AUDIOCPP_TASK_TTS = "tts"
 AUDIOCPP_TASK_VDES = "vdes"
 AUDIOCPP_SYNTHESIS_TASKS = (AUDIOCPP_TASK_TTS, "clon", AUDIOCPP_TASK_VDES)
 
+# The voice capability of a server model entry — how its voice is supplied.
+# Resolved per entry from (family, task, id) by
+# audiocpp_entry_voice_capability; drives both the Convert form (which
+# voice list to show) and the converter (speaker vs preset vs design mode).
+# Most families are clone-only; only the Qwen3-TTS CustomVoice entry has
+# built-in speakers, and only VoiceDesign entries take a description.
+AUDIOCPP_VOICE_SPEAKER = "speaker"  # built-in speaker name (Qwen CustomVoice)
+AUDIOCPP_VOICE_CLONE = "clone"      # server-side preset / voice_dir (Base, others)
+AUDIOCPP_VOICE_DESIGN = "design"    # voice described by --instructions (vdes)
+
 
 class AudioCppFamilyProfile:
-    """Request conventions of one audio.cpp model family."""
+    """Request conventions of one audio.cpp model family.
+
+    Language style and whether the family reads a style/instruction prompt;
+    these are family-level (every entry of a family shares them). Whether a
+    *specific entry* has built-in speakers is an entry-level concern, decided
+    by audiocpp_entry_voice_capability, not this profile.
+    """
 
     def __init__(self, language_style: str = AUDIOCPP_LANG_OMIT,
-                 sends_instructions: bool = False,
-                 builtin_speakers: bool = False):
+                 sends_instructions: bool = False):
         self.language_style = language_style
         self.sends_instructions = sends_instructions
-        self.builtin_speakers = builtin_speakers
 
 
 # Generic profile for families not listed in AUDIOCPP_FAMILY_PROFILES:
@@ -161,7 +180,6 @@ AUDIOCPP_FAMILY_PROFILES = {
     AUDIOCPP_FAMILY_QWEN3_TTS: AudioCppFamilyProfile(
         language_style=AUDIOCPP_LANG_DISPLAY,
         sends_instructions=True,
-        builtin_speakers=True,
     ),
     # Families whose language option takes a code (e.g. "en") instead of
     # a Qwen display name; otherwise clone-only like the default profile.
@@ -172,6 +190,34 @@ AUDIOCPP_FAMILY_PROFILES = {
     "supertonic": AudioCppFamilyProfile(language_style=AUDIOCPP_LANG_ISO),
 }
 
+
+def audiocpp_entry_voice_capability(family: str, task: str,
+                                    model_id: str) -> str:
+    """How a server model entry's voice is supplied — speaker/clone/design.
+
+    Resolved from the entry's family, task and id — the same {id, family,
+    task} triple GET /v1/models reports, so it works for local server.json
+    entries and remote live-queried entries alike. Qwen3-TTS is the one
+    family hosting several model *types* under one family id: the
+    CustomVoice model (id contains "customvoice") has built-in speakers, the
+    Base model and any other entry are clone-only, and VoiceDesign entries
+    (task "vdes") take a description. Every other family is clone-only.
+    """
+    if task == AUDIOCPP_TASK_VDES:
+        return AUDIOCPP_VOICE_DESIGN
+    if family == AUDIOCPP_FAMILY_QWEN3_TTS \
+            and "customvoice" in (model_id or "").lower():
+        return AUDIOCPP_VOICE_SPEAKER
+    return AUDIOCPP_VOICE_CLONE
+
+# Built-in CustomVoice speaker names for the Qwen3-TTS family. Shared by the
+# qwen-tts demo backend (config.SPEAKER, the qwen setup/form) and the
+# audio.cpp audiocpp backend's CustomVoice entry (the Convert form's Speaker
+# picker). Entries are the canonical/config form; speaker_display_name()
+# maps them to the wire (display) form via SPEAKER_DISPLAY_NAMES below.
+QWEN3_TTS_SPEAKERS = ("Vivian", "Serena", "Uncle_Fu", "Dylan", "Eric",
+                      "Ryan", "Aiden", "Ono_Anna", "Sohee")
+
 # Canonical speaker names -> display names used by the qwen-tts demo.
 SPEAKER_DISPLAY_NAMES = {
     "ryan": "Ryan",
@@ -209,10 +255,21 @@ def _resolve_request_seed() -> int:
     return seed
 
 
+def speaker_display_name_for(name: str) -> str:
+    """Return the wire (display) form of a Qwen3-TTS CustomVoice speaker NAME.
+
+    Accepts either the canonical/config form (e.g. "uncle_fu", "Uncle_Fu")
+    or the display form ("Uncle Fu"), case-insensitively; unknown names pass
+    through unchanged. Used by AudioCppTTSClient to normalize the --speaker /
+    Speaker-picker value into what audiocpp_server expects in the request's
+    voice field.
+    """
+    return SPEAKER_DISPLAY_NAMES.get((name or "").lower(), name)
+
+
 def speaker_display_name() -> str:
     """Return the display name for the configured custom speaker."""
-    return SPEAKER_DISPLAY_NAMES.get(
-        config.SPEAKER.lower(), config.SPEAKER)
+    return speaker_display_name_for(config.SPEAKER)
 
 
 def normalize_language(value: Optional[str]) -> str:
@@ -808,30 +865,39 @@ class AudioCppTTSClient(_BaseTTSClient):
     serving stack). The server API is family-agnostic; the family and task
     of the configured model entry are read from GET /v1/models at startup
     and adapt the request payload (language field style, style instructions)
-    through AUDIOCPP_FAMILY_PROFILES. Three voice modes, all resolved
-    server-side from the request's "voice"/"instructions" fields:
-
-    - Speaker mode (no ``voice``): Qwen3-TTS only. A built-in CustomVoice
-      speaker name (e.g. "Vivian") is passed through, plus the INSTRUCT
-      style prompt. The server must be configured with the CustomVoice
-      model for this. Families without built-in speakers reject this mode
-      with a hint to pass --voice (or --instructions, see below).
-    - Preset mode (``voice=NAME``): a voice configured on the server
+    through AUDIOCPP_FAMILY_PROFILES. The entry's voice capability
+    (audiocpp_entry_voice_capability: speaker / clone / design) decides how
+    its voice is supplied; all three are resolved server-side from the
+    request's "voice"/"instructions" fields:
+
+    - Speaker mode (--speaker NAME, or no flag on a CustomVoice entry):
+      Qwen3-TTS CustomVoice only. A built-in speaker name (e.g. "Vivian")
+      is passed through, plus the INSTRUCT style prompt. The selected entry
+      must be the CustomVoice model (capability == speaker); selecting it
+      on a non-speaker entry fails fast with a hint.
+    - Preset mode (--voice NAME): a voice configured on the server
       (``voice_presets`` or ``voice_dir`` in its config, e.g. a cloning
       reference). The name is validated against GET /v1/audio/voices at
       startup because an unresolvable name would silently fall back to
       plain TTS on a clone-based model instead of failing. When
       AUDIOCPP_CLONE_MODEL_ID names a second server entry of the same
       family (typically the Qwen Base model), preset requests are routed
-      to it. Only the entry actually used needs to exist on the server: a
-      clone-only (Base) server works for --voice runs, while speaker mode
-      on such a server fails with a hint to pass --voice.
+      to it. Selecting --voice on a CustomVoice primary with a clone id
+      configured is the documented way to switch a speaker setup to
+      cloning; without a clone id the voice is validated against the
+      server's voice library.
     - Voice design (task "vdes" entries, e.g. Qwen3-TTS VoiceDesign): the
       voice is described in natural language through ``instructions``,
       which is required and sent with every request (no ``voice`` field).
       A constant per-run seed keeps the designed voice consistent across
       chunk boundaries.
 
+    With neither --voice nor --speaker the entry's capability picks the
+    mode: design entries require --instructions; speaker entries use the
+    built-in CustomVoice speaker in config.SPEAKER; clone entries (the
+    Base model, and every other family) fail fast with a hint to pass
+    --voice, instead of silently synthesizing with a random default voice.
+
     ``instructions`` also works on non-design entries, where it acts as a
     generic style/delivery instruction (voice control): families that read
     it (OmniVoice, Qwen3-TTS CustomVoice, ...) shape the voice or delivery
@@ -852,7 +918,13 @@ class AudioCppTTSClient(_BaseTTSClient):
                  api_url: Optional[str] = None,
                  model_id: Optional[str] = None,
                  instructions: Optional[str] = None,
-                 request_options: Optional[Dict[str, str]] = None):
+                 request_options: Optional[Dict[str, str]] = None,
+                 speaker: Optional[str] = None):
+        if voice and speaker:
+            raise ValueError(
+                "--voice and --speaker are mutually exclusive: --voice selects "
+                "a server-side preset (cloning), --speaker a built-in "
+                "CustomVoice speaker (see README).")
         self.api_url = (api_url or config.AUDIOCPP_API_URL).rstrip("/")
         # Per-run model selection: the --model CLI flag overrides config; an
         # empty value is resolved at connect time when the server hosts exactly
@@ -868,8 +940,15 @@ class AudioCppTTSClient(_BaseTTSClient):
         # negative "randomize" seed, so a negative value means "send no seed
         # at all" (see _request_wav) and the server randomizes.
         self._seed = _resolve_request_seed()
+        # Voice/picker selection. preset_mode is True iff a server-side
+        # preset (--voice) was requested: it gates _select_model's reroute
+        # to AUDIOCPP_CLONE_MODEL_ID and the INSTRUCT style-prompt logic.
+        # speaker is a built-in CustomVoice speaker name; the entry's
+        # capability (resolved in _connect) validates it. The request's
+        # "voice" field (self.voice) is filled in _connect per the mode.
         self.preset_mode = bool(voice)
-        self.voice = voice or speaker_display_name()
+        self.speaker = speaker or None
+        self.voice = voice or None
         # Style/voice-design instruction sent with every request (the CLI
         # --instructions flag overrides AUDIOCPP_INSTRUCTIONS in config.py).
         # For task "vdes" entries it describes the voice to design; for other
@@ -879,9 +958,10 @@ class AudioCppTTSClient(_BaseTTSClient):
         # Free-form per-request options (--option KEY=VALUE) forwarded in the
         # request's "options" object; models ignore keys they don't know.
         self.request_options: Dict[str, str] = dict(request_options or {})
-        # Both set during _connect once the entry's task is known: design_mode
-        # for "vdes" entries, instruction_voice when a family without built-in
-        # speakers gets its voice from the instruction alone (no voice field).
+        # Set during _connect: design_mode for "vdes" entries, instruction_voice
+        # when a family without built-in speakers gets its voice from the
+        # instruction alone (no voice field). self.voice is also finalized
+        # there (the speaker/preset name, or config.SPEAKER for the default).
         self.design_mode = False
         self.instruction_voice = False
         # Family and task of the selected model entry and the family's request
@@ -898,12 +978,15 @@ class AudioCppTTSClient(_BaseTTSClient):
     def _connect(self) -> None:
         """Health-check the server and resolve the model, family, task, and voice.
 
-        Speaker mode is only offered to families with built-in speakers
-        (Qwen3-TTS); every other family must select a server-side voice
-        with --voice or describe one with --instructions, so it fails fast
-        with a hint instead of silently synthesizing with a random default
-        voice. Voice design entries (task "vdes") require --instructions
-        and reject --voice.
+        The entry's voice capability (audiocpp_entry_voice_capability, from
+        family/task/id) plus the caller's --voice/--speaker/--instructions
+        pick the mode: design entries require --instructions and reject
+        --voice/--speaker; a --speaker name requires a speaker-capable
+        (CustomVoice) entry; a --voice preset validates against the server's
+        voice library (and reroutes to AUDIOCPP_CLONE_MODEL_ID when set);
+        with neither flag, speaker-capable entries use the built-in
+        config.SPEAKER and clone entries fail fast with a hint instead of
+        silently synthesizing with a random default voice.
         """
         self._check_health()
         models = self._list_models()
@@ -921,11 +1004,13 @@ class AudioCppTTSClient(_BaseTTSClient):
                 f"model entries (tasks {', '.join(AUDIOCPP_SYNTHESIS_TASKS)}). "
                 f"Pick a synthesis entry with --model (available: {available})."
             )
+        capability = audiocpp_entry_voice_capability(
+            self.family, self.task, self.model_id)
         if self.design_mode:
-            if self.preset_mode:
+            if self.voice or self.speaker:
                 raise RuntimeError(
-                    f"--voice cannot be used with the voice design model "
-                    f"'{self.model_id}': the voice is described by the "
+                    f"--voice/--speaker cannot be used with the voice design "
+                    f"model '{self.model_id}': the voice is described by the "
                     "--instructions text instead (see README).")
             if not self.instructions:
                 raise RuntimeError(
@@ -938,18 +1023,31 @@ class AudioCppTTSClient(_BaseTTSClient):
                   f"(model '{self.model_id}', family '{self.family}', "
                   "voice design)")
             print(f"[INFO] Designing the voice from: {self.instructions}")
-        elif self.preset_mode:
+        elif self.speaker is not None:
+            if capability != AUDIOCPP_VOICE_SPEAKER:
+                raise RuntimeError(
+                    f"The audio.cpp model '{self.model_id}' (family "
+                    f"'{self.family}') has no built-in speakers; --speaker "
+                    "selects a CustomVoice speaker, so choose the "
+                    "CustomVoice model entry, or use --voice for a "
+                    "server-side preset (see README).")
+            self.voice = speaker_display_name_for(self.speaker)
+            print(f"[OK] Connected to audio.cpp server at {self.api_url} "
+                  f"(model '{self.model_id}', family '{self.family}', "
+                  f"speaker '{self.voice}')")
+        elif self.voice is not None:
+            # Explicit server-side preset (--voice): validate it and route
+            # to the clone model entry when AUDIOCPP_CLONE_MODEL_ID is set.
             self._check_voice()
             print(f"[OK] Connected to audio.cpp server at {self.api_url} "
                   f"(model '{self.model_id}', family '{self.family}', "
                   f"voice '{self.voice}')")
-        elif self.profile.builtin_speakers:
+        elif capability == AUDIOCPP_VOICE_SPEAKER:
+            # No flag on a CustomVoice entry: the built-in config.SPEAKER.
+            self.voice = speaker_display_name()
             print(f"[OK] Connected to audio.cpp server at {self.api_url} "
                   f"(model '{self.model_id}', family '{self.family}', "
                   f"speaker '{self.voice}')")
-            print("[INFO] Speaker mode expects the server to be configured with the "
-                  "CustomVoice model; with the Base model the speaker name is ignored "
-                  "and a random default voice is used (see README).")
         elif self.instructions:
             # Families without built-in speakers can still get their voice
             # from the instruction alone (e.g. OmniVoice voice design).
@@ -965,7 +1063,8 @@ class AudioCppTTSClient(_BaseTTSClient):
                 "must come from the server: rerun with --voice NAME "
                 "matching a voice_preset or voice_dir entry in the server "
                 "config, or describe a voice with --instructions for "
-                "families that support it (see README).")
+                "families that support it, or select the CustomVoice entry "
+                "for built-in speakers (see README).")
         if self.instructions and not self.design_mode and not self.instruction_voice:
             print(f"[INFO] Sending instruction with every request: {self.instructions}")
             print("[INFO] Its effect (style, emotion, delivery) depends on the "
@@ -1135,10 +1234,8 @@ class AudioCppTTSClient(_BaseTTSClient):
             # (presets are ignored on the CustomVoice model), so that case
             # keeps the warning; single-model servers of other families are
             # the normal configuration and only get a debug note.
-            primary_is_qwen = (families.get(self.model_id)
-                               or AUDIOCPP_FAMILY_QWEN3_TTS) \
-                == AUDIOCPP_FAMILY_QWEN3_TTS
-            if primary_is_qwen:
+            primary_family = families.get(self.model_id) or ""
+            if primary_family == AUDIOCPP_FAMILY_QWEN3_TTS:
                 logger.warning(
                     "AUDIOCPP_CLONE_MODEL_ID %r is not configured on the audio.cpp "
                     "server; preset requests use '%s' instead",
@@ -1166,22 +1263,21 @@ class AudioCppTTSClient(_BaseTTSClient):
     def _resolve_family(self, models: List[Dict[str, str]]) -> None:
         """Resolve the selected model's family and its request profile.
 
-        The family comes from GET /v1/models. Servers that predate the
-        family field served Qwen3-TTS only, so a missing family is treated
-        as qwen3_tts, which also preserves this client's legacy behavior
-        against those versions.
+        The family comes from GET /v1/models; a missing family is an unknown
+        family that falls through to the generic (clone-only) profile rather
+        than guessing a specific one — audiocpp_server always reports family
+        for entries its server.json describes.
         """
         entry = next(
             (model for model in models if model["id"] == self.model_id), None)
         family = (entry["family"] if entry is not None else "") or ""
-        if not family:
-            family = AUDIOCPP_FAMILY_QWEN3_TTS
-            logger.debug("Model '%s' reported no family; assuming qwen3_tts",
-                         self.model_id)
         self.family = family
         self.profile = AUDIOCPP_FAMILY_PROFILES.get(
             family, AUDIOCPP_DEFAULT_FAMILY_PROFILE)
-        if family not in AUDIOCPP_FAMILY_PROFILES:
+        if not family:
+            logger.debug("Model '%s' reported no family; using the generic "
+                         "profile", self.model_id)
+        elif family not in AUDIOCPP_FAMILY_PROFILES:
             logger.info(
                 "audio.cpp family '%s' has no dedicated profile; using the "
                 "generic profile (voice cloning via --voice, model-detected "
diff --git a/app/tests/test_converter.py b/app/tests/test_converter.py
index 9b0ddb4..317e772 100644
--- a/app/tests/test_converter.py
+++ b/app/tests/test_converter.py
@@ -131,6 +131,7 @@ class NarratorTagTests(unittest.TestCase):
         converter.voice_clone_ref_audio = ref_audio
         converter.backend = tts.BACKEND_QWEN
         converter.voice = None
+        converter.speaker = None
         converter.instructions = instructions
         return converter
 
@@ -159,11 +160,12 @@ class NarratorTagTests(unittest.TestCase):
         self.assertEqual(self._converter(tts.VOICE_MODE_CLONE, "/x/???.wav")._narrator_tag(),
                          "narrator")
 
-    def _audiocpp_converter(self, voice=None, instructions=None):
+    def _audiocpp_converter(self, voice=None, instructions=None, speaker=None):
         converter = self._converter(tts.VOICE_MODE_CUSTOM,
                                     instructions=instructions)
         converter.backend = tts.BACKEND_AUDIOCPP
         converter.voice = voice
+        converter.speaker = speaker
         return converter
 
     def test_audiocpp_design_run_uses_designed_tag(self):
@@ -181,6 +183,15 @@ class NarratorTagTests(unittest.TestCase):
         converter = self._audiocpp_converter()
         self.assertEqual(converter._narrator_tag(), "Vivian")
 
+    def test_audiocpp_explicit_speaker_uses_speaker_tag(self):
+        # A chosen CustomVoice speaker names the output, not config.SPEAKER.
+        converter = self._audiocpp_converter(speaker="Ryan")
+        self.assertEqual(converter._narrator_tag(), "Ryan")
+
+    def test_audiocpp_explicit_speaker_normalizes_display_name(self):
+        converter = self._audiocpp_converter(speaker="Uncle_Fu")
+        self.assertEqual(converter._narrator_tag(), "Uncle_Fu")
+
     def test_preflight_design_run_uses_designed_tag(self):
         with tempfile.TemporaryDirectory() as books_tmp, \
                 tempfile.TemporaryDirectory() as output_tmp:
@@ -547,6 +558,7 @@ class RunOverwritePromptTests(unittest.TestCase):
         self.converter.voice_clone_ref_audio = None
         self.converter.backend = tts.BACKEND_QWEN
         self.converter.voice = None
+        self.converter.speaker = None
         self.converter.instructions = None
         self.converter.speed = 1.0
         self.converter.single_file = False
diff --git a/app/tests/test_hub.py b/app/tests/test_hub.py
index 9842148..c5ca346 100644
--- a/app/tests/test_hub.py
+++ b/app/tests/test_hub.py
@@ -750,40 +750,75 @@ class ConvertFlowTests(unittest.TestCase):
                          [("audio.cpp [remote]", "audiocpp-remote")])
         # The model menu was fed from the live query (label, id).
         self.assertEqual(self._field("model_id")["choices"],
-                         [("higgs  (higgs_audio_tts, tts)", "higgs")])
+                         [("higgs  (higgs_audio_tts, clone)", "higgs")])
 
-    def test_audiocpp_qwen3_tts_voice_choices_lead_with_builtin_speaker(self):
+    def test_audiocpp_customvoice_entry_lists_builtin_speakers(self):
+        # A CustomVoice entry populates the Voice menu with the Qwen3-TTS
+        # built-in speakers and maps the pick to --speaker.
         self._patch_remote(
-            [{"id": "qwen", "family": "qwen3_tts", "task": "tts"}],
-            voices=["narrator"])
+            [{"id": "Qwen3-TTS-12Hz-1.7B-CustomVoice-GGUF",
+              "family": "qwen3_tts", "task": "tts"}])
         with patch.object(hub.config, "AUDIOCPP_INSTRUCTIONS", ""):
-            self._answer_form(backend="audiocpp-remote", model_id="qwen",
-                              audiocpp_voice="(built-in speaker)",
-                              instructions="")
+            self._answer_form(
+                backend="audiocpp-remote",
+                model_id="Qwen3-TTS-12Hz-1.7B-CustomVoice-GGUF",
+                audiocpp_voice="Ryan", instructions="")
             cmd = self._convert(
                 None, [self._remote("audiocpp", "audio.cpp")])
-        # The sentinel maps to "no voice" (built-in speaker).
+        # The speaker is passed as --speaker, not --voice.
         self.assertIsNone(cmd[2]["voice"])
+        self.assertEqual(cmd[2]["speaker"], "Ryan")
         fields = self.tui.forms_seen[0][1]
         voice_field = self._field("audiocpp_voice")
-        choices = voice_field["choices"](fields)
-        self.assertEqual(choices,
-                         [("(built-in speaker)", "(built-in speaker)"),
-                          ("narrator", "narrator")])
-
-    def test_audiocpp_remote_missing_family_treated_as_qwen3_tts(self):
-        # Legacy servers omit family/task; the converter defaults them to
-        # qwen3_tts/tts and so must the form (voice optional).
+        self.assertEqual(voice_field["choices"](fields),
+                         [(s, s) for s in hub.QWEN3_TTS_SPEAKERS])
+        # CustomVoice reads a style instruction, so the field stays visible.
+        instr = self._field("instructions")
+        self.assertTrue(instr["visible"](fields))
+        self.assertIsNone(instr["validate"](""))
+
+    def test_audiocpp_qwen3_tts_base_entry_lists_clone_voices(self):
+        # A Base entry populates the Voice menu with the server's clone
+        # voices only (no built-in speakers) and maps the pick to --voice.
+        self._patch_remote(
+            [{"id": "Qwen3-TTS-12Hz-1.7B-Base-GGUF",
+              "family": "qwen3_tts", "task": "tts"}],
+            voices=["narrator"])
+        with patch.object(hub.config, "AUDIOCPP_INSTRUCTIONS", ""):
+            self._answer_form(
+                backend="audiocpp-remote",
+                model_id="Qwen3-TTS-12Hz-1.7B-Base-GGUF",
+                audiocpp_voice="narrator", instructions="")
+            cmd = self._convert(
+                None, [self._remote("audiocpp", "audio.cpp")])
+        self.assertEqual(cmd[2]["voice"], "narrator")
+        self.assertIsNone(cmd[2]["speaker"])
+        self.assertIsNone(cmd[2]["instructions"])
+        fields = self.tui.forms_seen[0][1]
+        voice_field = self._field("audiocpp_voice")
+        self.assertEqual(voice_field["choices"](fields),
+                         [("narrator", "narrator")])
+        # Base (clone) ignores instructions, so the field is hidden.
+        instr = self._field("instructions")
+        self.assertFalse(instr["visible"](fields))
+
+    def test_audiocpp_remote_missing_family_is_clone_capable(self):
+        # A missing family is unknown — not guessed as qwen3_tts — so the
+        # entry is clone-only: it needs a --voice rather than offering a
+        # built-in speaker.
         self._patch_remote([{"id": "legacy", "family": "", "task": ""}],
                            voices=[])
         with patch.object(hub.config, "AUDIOCPP_INSTRUCTIONS", ""):
             self._answer_form(backend="audiocpp-remote", model_id="legacy",
-                              audiocpp_voice="(built-in speaker)",
-                              instructions="")
+                              audiocpp_voice="", instructions="")
             cmd = self._convert(
                 None, [self._remote("audiocpp", "audio.cpp")])
         self.assertIsNotNone(cmd)
         self.assertIsNone(cmd[2]["voice"])
+        self.assertIsNone(cmd[2]["speaker"])
+        voice_field = self._field("audiocpp_voice")
+        # Clone-only: an empty voice is refused (no built-in speaker option).
+        self.assertIsNotNone(voice_field["validate"](""))
 
     def test_audiocpp_vdes_hides_voice_and_requires_instructions(self):
         self._patch_remote(
@@ -800,9 +835,24 @@ class ConvertFlowTests(unittest.TestCase):
         voice_field = self._field("audiocpp_voice")
         self.assertFalse(voice_field["visible"](fields))
         instr = self._field("instructions")
+        self.assertTrue(instr["visible"](fields))
         self.assertIsNotNone(instr["validate"](""))
         self.assertIsNone(instr["validate"]("describe me"))
 
+    def test_audiocpp_clone_drops_stale_instructions(self):
+        # A Base/clone entry ignores instructions: even if the form held a
+        # leftover value, the mapper must not send it to the model.
+        self._patch_remote(
+            [{"id": "higgs", "family": "higgs_audio_tts", "task": "tts"}],
+            voices=["narrator"])
+        with patch.object(hub.config, "AUDIOCPP_INSTRUCTIONS", ""):
+            self._answer_form(backend="audiocpp-remote", model_id="higgs",
+                              audiocpp_voice="narrator",
+                              instructions="stale description")
+            cmd = self._convert(
+                None, [self._remote("audiocpp", "audio.cpp")])
+        self.assertIsNone(cmd[2]["instructions"])
+
     def test_audiocpp_required_voice_validates(self):
         # A non-qwen3_tts family needs a --voice; a blank value refuses.
         self._patch_remote(
@@ -1082,9 +1132,11 @@ class ConvertFlowTests(unittest.TestCase):
              "mode", "speaker", "clone", "output_format", "speed",
              "single_file", "debug"])
         # The form opens on the configured default (audio.cpp): its fields
-        # show, the other backend's hide.
-        for key in ("model_id", "audiocpp_voice", "instructions"):
+        # show, the other backend's hide. (Instructions is hidden too: the
+        # default higgs entry is clone-only, which ignores instructions.)
+        for key in ("model_id", "audiocpp_voice"):
             self.assertTrue(self._field(key)["visible"](fields))
+        self.assertFalse(self._field("instructions")["visible"](fields))
         for key in ("mode", "speaker", "clone"):
             self.assertFalse(self._field(key)["visible"](fields))
         # Picking qwen in the Backend field swaps which options show.
@@ -1100,8 +1152,9 @@ class ConvertFlowTests(unittest.TestCase):
             self.assertFalse(self._field(key)["visible"](fields))
         # And back to audio.cpp.
         fields[0]["value"] = "audiocpp"
-        for key in ("model_id", "audiocpp_voice", "instructions"):
+        for key in ("model_id", "audiocpp_voice"):
             self.assertTrue(self._field(key)["visible"](fields))
+        self.assertFalse(self._field("instructions")["visible"](fields))
         for key in ("mode", "speaker", "clone"):
             self.assertFalse(self._field(key)["visible"](fields))
 
diff --git a/app/tests/test_tts.py b/app/tests/test_tts.py
index b43919d..0e35f78 100644
--- a/app/tests/test_tts.py
+++ b/app/tests/test_tts.py
@@ -479,8 +479,10 @@ class AudioCppTTSClientHealthTests(unittest.TestCase):
 
     def setUp(self):
         # The default AUDIOCPP_MODEL_ID is empty (auto-select); these tests
-        # exercise a configured single-model server, so pin a concrete id.
-        patcher = patch.object(config, "AUDIOCPP_MODEL_ID", "qwen")
+        # exercise a configured single-model CustomVoice server, so pin a
+        # concrete id whose "customvoice" substring marks it speaker-capable.
+        patcher = patch.object(
+            config, "AUDIOCPP_MODEL_ID", "Qwen3-TTS-12Hz-1.7B-CustomVoice-GGUF")
         patcher.start()
         self.addCleanup(patcher.stop)
 
@@ -500,7 +502,8 @@ class AudioCppTTSClientHealthTests(unittest.TestCase):
                                            else {"status": "ok"})
             if url.endswith("/v1/models"):
                 return self._json_response(models if models is not None else
-                                           {"data": [{"id": config.AUDIOCPP_MODEL_ID}]})
+                                           {"data": [{"id": config.AUDIOCPP_MODEL_ID,
+                                                      "family": "qwen3_tts"}]})
             if "/v1/audio/voices" in url:
                 if voices is Exception:
                     raise Exception("voices endpoint down")
@@ -509,11 +512,12 @@ class AudioCppTTSClientHealthTests(unittest.TestCase):
             raise AssertionError(f"unexpected URL: {url}")
         return _dispatch
 
-    def _client(self, voice=None, language=None, model_id=None, **kwargs):
+    def _client(self, voice=None, language=None, model_id=None, speaker=None,
+                **kwargs):
         with patch("converter.tts.urllib.request.urlopen",
                    side_effect=self._get_responses(**kwargs)):
             return AudioCppTTSClient(voice=voice, language=language,
-                                     model_id=model_id)
+                                     model_id=model_id, speaker=speaker)
 
     def test_unreachable_server_raises_with_readme_pointer(self):
         import urllib.error
@@ -551,6 +555,42 @@ class AudioCppTTSClientHealthTests(unittest.TestCase):
             client = self._client()
         self.assertEqual(client.voice, "Uncle Fu")
 
+    def test_explicit_speaker_selects_speaker_mode(self):
+        # --speaker picks a CustomVoice speaker; the name is normalized to
+        # its wire (display) form and no preset validation runs.
+        client = self._client(speaker="Uncle_Fu")
+        self.assertEqual(client.voice, "Uncle Fu")
+        self.assertFalse(client.preset_mode)
+
+    def test_explicit_speaker_on_clone_entry_raises(self):
+        # --speaker is meaningless on a clone-only (Base) entry.
+        with self.assertRaises(RuntimeError) as ctx:
+            self._client(speaker="Ryan", model_id="Qwen3-TTS-12Hz-1.7B-Base-GGUF",
+                         models={"data": [
+                             {"id": "Qwen3-TTS-12Hz-1.7B-Base-GGUF",
+                              "family": "qwen3_tts"}]})
+        message = str(ctx.exception)
+        self.assertIn("no built-in speakers", message)
+        self.assertIn("--voice", message)
+
+    def test_voice_and_speaker_are_mutually_exclusive(self):
+        with self.assertRaises(ValueError) as ctx:
+            AudioCppTTSClient(voice="narrator", speaker="Ryan")
+        self.assertIn("mutually exclusive", str(ctx.exception))
+
+    def test_no_voice_on_base_entry_raises_instead_of_silent_speaker(self):
+        # The Base model has no built-in speakers: without --voice the run
+        # fails fast instead of silently sending a speaker name that the
+        # model ignores.
+        with self.assertRaises(RuntimeError) as ctx:
+            self._client(model_id="Qwen3-TTS-12Hz-1.7B-Base-GGUF",
+                         models={"data": [
+                             {"id": "Qwen3-TTS-12Hz-1.7B-Base-GGUF",
+                              "family": "qwen3_tts"}]})
+        message = str(ctx.exception)
+        self.assertIn("Base-GGUF", message)
+        self.assertIn("--voice", message)
+
     def test_preset_mode_uses_requested_voice(self):
         client = self._client(voice="narrator")
         self.assertEqual(client.voice, "narrator")
@@ -598,7 +638,8 @@ class AudioCppTTSClientHealthTests(unittest.TestCase):
                 self.assertLogs("converter.tts", level="WARNING") as logs:
             client = self._client(
                 voice="narrator",
-                models={"data": [{"id": "qwen3-tts"}, {"id": "pocket-tts"}]})
+                models={"data": [{"id": "qwen3-tts", "family": "qwen3_tts"},
+                                 {"id": "pocket-tts"}]})
         self.assertEqual(client.model_id, "qwen3-tts")
         self.assertTrue(any("qwen3-tts-clone" in line for line in logs.output))
 
@@ -632,11 +673,17 @@ class AudioCppTTSClientHealthTests(unittest.TestCase):
         self.assertEqual(client.model_id, "higgs")
 
     def test_clone_model_id_ignored_for_speaker_mode(self):
-        with patch.object(config, "AUDIOCPP_MODEL_ID", "qwen3-tts"), \
+        # Speaker mode (no --voice on a CustomVoice entry) never reroutes to
+        # AUDIOCPP_CLONE_MODEL_ID — that reroute is a preset-mode concern.
+        with patch.object(config, "AUDIOCPP_MODEL_ID",
+                          "Qwen3-TTS-CustomVoice"), \
                 patch.object(config, "AUDIOCPP_CLONE_MODEL_ID", "qwen3-tts-clone"):
             client = self._client(
-                models={"data": [{"id": "qwen3-tts"}, {"id": "qwen3-tts-clone"}]})
-        self.assertEqual(client.model_id, "qwen3-tts")
+                models={"data": [{"id": "Qwen3-TTS-CustomVoice",
+                                  "family": "qwen3_tts"},
+                                 {"id": "qwen3-tts-clone",
+                                  "family": "qwen3_tts"}]})
+        self.assertEqual(client.model_id, "Qwen3-TTS-CustomVoice")
 
     def test_clone_model_id_equal_to_primary_is_noop(self):
         with patch.object(config, "AUDIOCPP_CLONE_MODEL_ID",
@@ -662,9 +709,12 @@ class AudioCppTTSClientHealthTests(unittest.TestCase):
         self.assertIn("--voice", message)
 
     def test_preset_mode_with_no_matching_model_lists_both_ids(self):
+        # Neither the primary nor the clone id is on the server, so the
+        # family is unknown and no degradation warning is logged — the
+        # requirement error lists both configured ids instead.
         with patch.object(config, "AUDIOCPP_MODEL_ID", "qwen3-tts"), \
                 patch.object(config, "AUDIOCPP_CLONE_MODEL_ID", "qwen3-tts-clone"), \
-                self.assertLogs("converter.tts", level="WARNING"):
+                self.assertNoLogs("converter.tts", level="WARNING"):
             with self.assertRaises(RuntimeError) as ctx:
                 self._client(voice="narrator",
                              models={"data": [{"id": "pocket-tts"}]})
@@ -678,7 +728,10 @@ class AudioCppTaskDetectionTests(unittest.TestCase):
     """Task auto-detection (tts/clon/vdes) and voice design validation."""
 
     def setUp(self):
-        patcher = patch.object(config, "AUDIOCPP_MODEL_ID", "qwen")
+        # Pin a CustomVoice id so the default (no-voice) path is speaker
+        # mode; individual tests override family/task to exercise other paths.
+        patcher = patch.object(
+            config, "AUDIOCPP_MODEL_ID", "Qwen3-TTS-12Hz-1.7B-CustomVoice-GGUF")
         patcher.start()
         self.addCleanup(patcher.stop)
 
@@ -853,18 +906,19 @@ class AudioCppFamilyDetectionTests(unittest.TestCase):
         self.assertEqual(client.family, "higgs_audio_tts")
         self.assertIs(client.profile, tts.AUDIOCPP_DEFAULT_FAMILY_PROFILE)
 
-    def test_missing_family_falls_back_to_qwen3_tts(self):
+    def test_missing_family_uses_generic_profile(self):
+        # A missing family is unknown (not guessed as qwen3_tts): it falls
+        # through to the generic clone-only profile.
         client = self._client(models={"data": [
             {"id": config.AUDIOCPP_MODEL_ID}]})
-        self.assertEqual(client.family, "qwen3_tts")
-        self.assertTrue(client.profile.builtin_speakers)
+        self.assertEqual(client.family, "")
+        self.assertIs(client.profile, tts.AUDIOCPP_DEFAULT_FAMILY_PROFILE)
 
     def test_unknown_family_uses_generic_profile(self):
         client = self._client(models={"data": [
             {"id": config.AUDIOCPP_MODEL_ID, "family": "future_tts"}]})
         self.assertEqual(client.family, "future_tts")
         self.assertIs(client.profile, tts.AUDIOCPP_DEFAULT_FAMILY_PROFILE)
-        self.assertFalse(client.profile.builtin_speakers)
         self.assertEqual(client.profile.language_style, tts.AUDIOCPP_LANG_OMIT)
 
     def test_speaker_mode_rejected_for_clone_only_family(self):
@@ -879,11 +933,31 @@ class AudioCppFamilyDetectionTests(unittest.TestCase):
             self.assertIn("no built-in speakers", message)
         self.assertIsNone(client)
 
-    def test_speaker_mode_allowed_for_qwen_family(self):
-        client = self._client(voice=None, models={"data": [
-            {"id": config.AUDIOCPP_MODEL_ID, "family": "qwen3_tts"}]})
+    def test_speaker_mode_allowed_for_customvoice_entry(self):
+        # A Qwen3-TTS entry whose id names CustomVoice is speaker-capable;
+        # no --voice is needed.
+        with patch.object(config, "AUDIOCPP_MODEL_ID",
+                          "Qwen3-TTS-12Hz-1.7B-CustomVoice-GGUF"):
+            client = self._client(voice=None, models={"data": [
+                {"id": "Qwen3-TTS-12Hz-1.7B-CustomVoice-GGUF",
+                 "family": "qwen3_tts"}]})
         self.assertEqual(client.family, "qwen3_tts")
 
+    def test_speaker_mode_rejected_for_qwen_base_entry(self):
+        # A Qwen3-TTS entry whose id names Base (not CustomVoice) is
+        # clone-only, even though its family has built-in speakers on other
+        # entries: without --voice it fails fast.
+        client = None
+        try:
+            client = self._client(voice=None, models={"data": [
+                {"id": "Qwen3-TTS-12Hz-1.7B-Base-GGUF",
+                 "family": "qwen3_tts"}]})
+        except RuntimeError as exc:
+            message = str(exc)
+            self.assertIn("Base-GGUF", message)
+            self.assertIn("--voice", message)
+        self.assertIsNone(client)
+
     def test_clone_model_id_of_different_family_is_ignored(self):
         with patch.object(config, "AUDIOCPP_MODEL_ID", "higgs"), \
                 patch.object(config, "AUDIOCPP_CLONE_MODEL_ID", "qwen-clone"), \
@@ -919,6 +993,52 @@ class AudioCppFamilyDetectionTests(unittest.TestCase):
         self.assertIsNone(tts.LANGUAGE_ISO_CODES.get("Auto"))
 
 
+class AudiocppEntryVoiceCapabilityTests(unittest.TestCase):
+    """The per-entry voice capability resolver (speaker/clone/design)."""
+
+    def _cap(self, family="", task="tts", model_id=""):
+        return tts.audiocpp_entry_voice_capability(family, task, model_id)
+
+    def test_vdes_task_is_design(self):
+        self.assertEqual(self._cap("qwen3_tts", "vdes",
+                                   "Qwen3-TTS-VoiceDesign-GGUF"),
+                         tts.AUDIOCPP_VOICE_DESIGN)
+
+    def test_qwen_customvoice_entry_is_speaker(self):
+        self.assertEqual(self._cap("qwen3_tts", "tts",
+                                   "Qwen3-TTS-12Hz-1.7B-CustomVoice-GGUF"),
+                         tts.AUDIOCPP_VOICE_SPEAKER)
+
+    def test_qwen_base_entry_is_clone(self):
+        self.assertEqual(self._cap("qwen3_tts", "tts",
+                                   "Qwen3-TTS-12Hz-1.7B-Base-GGUF"),
+                         tts.AUDIOCPP_VOICE_CLONE)
+
+    def test_qwen_unidentified_entry_is_clone(self):
+        self.assertEqual(self._cap("qwen3_tts", "tts", "qwen"),
+                         tts.AUDIOCPP_VOICE_CLONE)
+
+    def test_other_families_are_clone(self):
+        self.assertEqual(self._cap("higgs_audio_tts", "tts", "higgs"),
+                         tts.AUDIOCPP_VOICE_CLONE)
+
+    def test_missing_family_is_clone(self):
+        self.assertEqual(self._cap("", "tts", "legacy"),
+                         tts.AUDIOCPP_VOICE_CLONE)
+
+    def test_customvoice_match_is_case_insensitive(self):
+        self.assertEqual(self._cap("qwen3_tts", "tts",
+                                   "Qwen3-TTS-12Hz-1.7B-CUSTOMVOICE-GGUF"),
+                         tts.AUDIOCPP_VOICE_SPEAKER)
+
+    def test_customvoice_id_in_other_family_is_not_speaker(self):
+        # The "customvoice" substring only marks a speaker for the qwen3_tts
+        # family; another family with a lookalike id stays clone-only.
+        self.assertEqual(self._cap("future_tts", "tts",
+                                   "Qwen3-TTS-CustomVoice"),
+                         tts.AUDIOCPP_VOICE_CLONE)
+
+
 class AudioCppTTSClientRequestTests(unittest.TestCase):
     """The /v1/audio/speech payload and response validation."""
 
@@ -943,6 +1063,7 @@ class AudioCppTTSClientRequestTests(unittest.TestCase):
         client.model_id = config.AUDIOCPP_MODEL_ID
         client.preset_mode = preset_mode
         client.voice = voice
+        client.speaker = None
         client.language = language
         client._seed = seed
         client.family = family
@@ -953,10 +1074,12 @@ class AudioCppTTSClientRequestTests(unittest.TestCase):
         client.request_options = dict(request_options or {})
         client.design_mode = task == tts.AUDIOCPP_TASK_VDES
         # Mirrors the connect-time rule: an instruction-defined voice on a
-        # family without built-in speakers (design mode takes precedence).
+        # clone-capable entry with no --voice (design mode takes precedence).
+        capability = tts.audiocpp_entry_voice_capability(
+            family, task, client.model_id)
         client.instruction_voice = (
             not preset_mode and not client.design_mode
-            and not client.profile.builtin_speakers
+            and capability == tts.AUDIOCPP_VOICE_CLONE
             and bool(client.instructions))
         return client
 
@@ -1396,6 +1519,7 @@ class AudioCppUnloadModelsTests(unittest.TestCase):
         client.design_mode = False
         client.instruction_voice = False
         client.instructions = ""
+        client.speaker = None
         with patch.object(client, "_check_health"), \
                 patch.object(client, "_list_models",
                              return_value=[{"id": client.model_id,
@@ -1425,6 +1549,7 @@ class AudioCppUnloadModelsTests(unittest.TestCase):
         client.design_mode = False
         client.instruction_voice = False
         client.instructions = ""
+        client.speaker = None
         with patch.object(client, "_check_health"), \
                 patch.object(client, "_list_models",
                              return_value=[{"id": client.model_id,
@@ -1466,7 +1591,8 @@ class BackendWiringTests(unittest.TestCase):
                                               model_id=None,
                                               instructions=None,
                                               request_options={},
-                                              api_url=None)
+                                              api_url=None,
+                                              speaker=None)
         mock_faster.assert_not_called()
         mock_qwen.assert_not_called()
 
@@ -1478,7 +1604,8 @@ class BackendWiringTests(unittest.TestCase):
                                               model_id=None,
                                               instructions=None,
                                               request_options={},
-                                              api_url=None)
+                                              api_url=None,
+                                              speaker=None)
 
     def test_audiocpp_backend_model_id_is_wired_through(self):
         with patch("converter.converter.AudioCppTTSClient") as mock_audiocpp:
@@ -1488,7 +1615,7 @@ class BackendWiringTests(unittest.TestCase):
         mock_audiocpp.assert_called_once_with(
             voice="narrator", language=config.LANGUAGE,
             model_id="higgs", instructions=None,
-            request_options={}, api_url=None)
+            request_options={}, api_url=None, speaker=None)
 
     def test_audiocpp_backend_instructions_and_options_are_wired_through(self):
         with patch("converter.converter.AudioCppTTSClient") as mock_audiocpp:
@@ -1502,7 +1629,7 @@ class BackendWiringTests(unittest.TestCase):
             model_id=None,
             instructions="A warm adult narrator",
             request_options={"emotion": "neutral", "speed": "1.1"},
-            api_url=None)
+            api_url=None, speaker=None)
 
     def test_qwen_backend_uses_qwen_client(self):
         with patch("converter.converter.FasterTTSClient") as mock_faster, \
@@ -1529,7 +1656,7 @@ class BackendWiringTests(unittest.TestCase):
         mock_audiocpp.assert_called_once_with(
             voice="narrator", language=config.LANGUAGE, model_id=None,
             instructions=None, request_options={},
-            api_url="http://10.0.0.5:8080")
+            api_url="http://10.0.0.5:8080", speaker=None)
         with patch("converter.converter.FasterTTSClient") as mock_faster:
             AudiobookConverter(voice_mode=tts.VOICE_MODE_CLONE,
                                backend=tts.BACKEND_FASTER, voice="narrator",
diff --git a/app/tests/test_tui.py b/app/tests/test_tui.py
index 964f983..543194e 100644
--- a/app/tests/test_tui.py
+++ b/app/tests/test_tui.py
@@ -204,6 +204,17 @@ class MenuTests(TuiTestCase):
         self.assertGreater(help_x, margin)
         self.assert_inside_border(screen)
 
+    def test_blank_line_below_the_title(self):
+        # A titled dialog reserves a blank line between the title row
+        # and the first body row.
+        screen = FakeScreen(keys=[10])
+        tui.menu(screen, "Pick", self.OPTIONS)
+        title_y = next(y for y, _, text, _ in screen.strings
+                       if text == " Pick ")
+        first_y = next(y for y, _, text, _ in screen.strings
+                       if text == "first option")
+        self.assertEqual(first_y, title_y + 2)
+
     def test_up_wraps_around_to_last_option(self):
         screen = FakeScreen(keys=[FakeCurses.KEY_UP, 10])
         value = tui.menu(screen, "Pick one", self.OPTIONS)
diff --git a/app/ui/hub.py b/app/ui/hub.py
index 29edf9b..84b4155 100644
--- a/app/ui/hub.py
+++ b/app/ui/hub.py
@@ -53,10 +53,14 @@ from converter.converter import (
     voice_mode_for,
 )
 from converter.tts import (
-    AUDIOCPP_FAMILY_QWEN3_TTS,
+    AUDIOCPP_VOICE_CLONE,
+    AUDIOCPP_VOICE_DESIGN,
+    AUDIOCPP_VOICE_SPEAKER,
     BACKEND_AUDIOCPP,
     BACKEND_FASTER,
     BACKEND_QWEN,
+    QWEN3_TTS_SPEAKERS,
+    audiocpp_entry_voice_capability,
     normalize_language,
 )
 from ui import runview, taskview, tui
@@ -737,6 +741,7 @@ def _preflight(stdscr, cmd: tuple) -> bool:
             voice_clone_ref_audio=kwargs.get("clone"),
             output_format=kwargs.get("output_format") or config.AUDIO_FORMAT,
             instructions=kwargs.get("instructions"),
+            speaker=kwargs.get("speaker"),
             confirm=confirm)
     if not book_files:
         tui.flash(stdscr, "No books to convert. Add a .txt, .pdf or .epub "
@@ -770,11 +775,6 @@ def _gate_backend(field: dict, key: str) -> Callable:
     return visible
 
 
-# Sentinel value the audio.cpp Voice field uses for "no --voice" (the
-# built-in CustomVoice speaker); mapped to None when the form returns.
-_AUDIOCPP_BUILTIN_SPEAKER = "(built-in speaker)"
-
-
 def _field_value(fields, key: str, default=None):
     """Current value of the field named KEY, or DEFAULT when absent."""
     for field in fields:
@@ -870,14 +870,13 @@ def _audiocpp_fields(stdscr, api_url: Optional[str] = None) -> Optional[tuple]:
             return None
 
     # Normalize each entry so the form logic sees a family/task always.
+    # A missing family is left empty (an unknown family resolves to the
+    # clone capability, requiring a --voice) rather than guessing a specific
+    # one — audiocpp_server always reports family for entries it hosts.
     models = [dict(m) for m in models]
     for entry in models:
         entry["family"] = entry.get("family") or ""
         entry["task"] = entry.get("task") or "tts"
-        if not local and not entry["family"]:
-            # Servers predating the family field omit it; mirror the
-            # converter's default: unknown family means qwen3_tts.
-            entry["family"] = AUDIOCPP_FAMILY_QWEN3_TTS
 
     if local:
         # Only offer entries whose model files are actually on disk: a
@@ -914,75 +913,100 @@ def _audiocpp_fields(stdscr, api_url: Optional[str] = None) -> Optional[tuple]:
         return next((m for m in models if m.get("id") == model_id),
                     models[0])
 
-    def model_task(fields) -> str:
-        return model_entry(fields).get("task", "tts")
-
-    def model_family(fields) -> str:
-        return model_entry(fields).get("family") or ""
+    def model_capability(fields) -> str:
+        entry = model_entry(fields)
+        return audiocpp_entry_voice_capability(
+            entry.get("family") or "", entry.get("task") or "tts",
+            entry.get("id") or "")
 
     def reset_voice(fields) -> None:
         """Re-point the Voice field at the newly selected model's voice."""
         voice_field = next(f for f in fields
                            if f.get("key") == "audiocpp_voice")
-        if model_task(fields) == "vdes":
+        capability = model_capability(fields)
+        if capability == AUDIOCPP_VOICE_DESIGN:
             voice_field["value"] = None
-        elif model_family(fields) == AUDIOCPP_FAMILY_QWEN3_TTS:
-            voice_field["value"] = _AUDIOCPP_BUILTIN_SPEAKER
-        else:
+        elif capability == AUDIOCPP_VOICE_SPEAKER:
+            voice_field["value"] = (config.SPEAKER
+                                    if config.SPEAKER in QWEN3_TTS_SPEAKERS
+                                    else QWEN3_TTS_SPEAKERS[0])
+        else:  # clone
             voices = voices_for(_field_value(fields, "model_id"))
             voice_field["value"] = voices[0] if voices else ""
 
     def voice_choices(fields) -> list:
-        if model_family(fields) == AUDIOCPP_FAMILY_QWEN3_TTS:
-            return [(_AUDIOCPP_BUILTIN_SPEAKER, _AUDIOCPP_BUILTIN_SPEAKER)] \
-                + [(v, v) for v in voices_for(_field_value(fields,
+        capability = model_capability(fields)
+        if capability == AUDIOCPP_VOICE_SPEAKER:
+            # Built-in Qwen3-TTS CustomVoice speakers; no server query needed.
+            return [(s, s) for s in QWEN3_TTS_SPEAKERS]
+        if capability == AUDIOCPP_VOICE_CLONE:
+            return [(v, v) for v in voices_for(_field_value(fields,
                                                            "model_id"))]
-        return [(v, v) for v in voices_for(_field_value(fields, "model_id"))]
+        return []  # design: the field is hidden
 
     model_ids = [m.get("id") for m in models]
     default_model = config.AUDIOCPP_MODEL_ID \
         if config.AUDIOCPP_MODEL_ID in model_ids else model_ids[0]
     default_entry = next((m for m in models if m.get("id") == default_model),
                          models[0])
-    initial_voice = _AUDIOCPP_BUILTIN_SPEAKER
-    if default_entry.get("task") == "vdes":
-        initial_voice = None
-    elif default_entry.get("family") != AUDIOCPP_FAMILY_QWEN3_TTS:
+    default_capability = audiocpp_entry_voice_capability(
+        default_entry.get("family") or "", default_entry.get("task") or "tts",
+        default_entry.get("id") or "")
+    initial_voice = None
+    if default_capability == AUDIOCPP_VOICE_SPEAKER:
+        initial_voice = (config.SPEAKER if config.SPEAKER in QWEN3_TTS_SPEAKERS
+                         else QWEN3_TTS_SPEAKERS[0])
+    elif default_capability == AUDIOCPP_VOICE_CLONE:
         initial = voices_for(default_model)
         initial_voice = initial[0] if initial else ""
 
+    def _label(entry: dict) -> str:
+        capability = audiocpp_entry_voice_capability(
+            entry.get("family") or "", entry.get("task") or "tts",
+            entry.get("id") or "")
+        return (f"{entry.get('id')}  ({entry.get('family') or '?'}, "
+                f"{capability})")
+
     fields = [
         {"key": "model_id", "label": "Model", "kind": "choice",
          "value": default_model,
-         "choices": [(f"{m.get('id')}  ({m.get('family') or '?'}, "
-                      f"{m.get('task') or 'tts'})", m.get("id"))
-                     for m in models],
+         "choices": [(_label(m), m.get("id")) for m in models],
          "on_change": reset_voice},
         {"key": "audiocpp_voice", "label": "Voice", "kind": "choice",
          "value": initial_voice,
          "choices": lambda fs: voice_choices(fs),
-         "visible": lambda fs: model_task(fs) != "vdes",
+         "visible": lambda fs: model_capability(fs) != AUDIOCPP_VOICE_DESIGN,
          "validate": lambda value: None
-         if (model_family(fields) == AUDIOCPP_FAMILY_QWEN3_TTS or value)
+         if (model_capability(fields) != AUDIOCPP_VOICE_CLONE or value)
          else "This model needs a voice — pick one or switch models"},
         {"key": "instructions", "label": "Instructions", "kind": "text",
          "value": config.AUDIOCPP_INSTRUCTIONS,
+         "visible": lambda fs: model_capability(fs) in (AUDIOCPP_VOICE_DESIGN,
+                                                        AUDIOCPP_VOICE_SPEAKER),
          "validate": lambda value: None
-         if (model_task(fields) != "vdes" or str(value).strip())
+         if (model_capability(fields) != AUDIOCPP_VOICE_DESIGN or str(value).strip())
          else "Describe the voice, e.g. 'A warm female narrator'"},
     ]
 
     def mapper(result) -> Optional[tuple]:
         model_id = result["model_id"]
-        voice = result["audiocpp_voice"]
-        if voice == _AUDIOCPP_BUILTIN_SPEAKER or not voice:
-            voice = None
         entry = next((m for m in models if m.get("id") == model_id), {})
-        if entry.get("task") == "vdes":
-            voice = None
-        instructions = (result["instructions"] or "").strip() or None
+        capability = audiocpp_entry_voice_capability(
+            entry.get("family") or "", entry.get("task") or "tts",
+            entry.get("id") or "")
+        picked = result["audiocpp_voice"]
+        voice = None
+        speaker = None
+        if capability == AUDIOCPP_VOICE_SPEAKER:
+            speaker = picked or None
+        elif capability == AUDIOCPP_VOICE_CLONE:
+            voice = picked or None
+        # design: neither — the voice comes from --instructions
+        instructions = None
+        if capability in (AUDIOCPP_VOICE_DESIGN, AUDIOCPP_VOICE_SPEAKER):
+            instructions = (result["instructions"] or "").strip() or None
         kwargs = {
-            "model_id": model_id, "voice": voice,
+            "model_id": model_id, "voice": voice, "speaker": speaker,
             "instructions": instructions,
             **_common_kwargs(result),
         }
diff --git a/app/ui/tui.py b/app/ui/tui.py
index 663df0a..d675946 100644
--- a/app/ui/tui.py
+++ b/app/ui/tui.py
@@ -373,7 +373,10 @@ class Frame:
         Returns (y0, x0, dialog_h, visible); also refreshes
         self.scroll and self.page_size.
         """
-        chrome = 7 if self.buttons else 6  # title/gap/status/footer/borders
+        # Borders, title, status and footer are fixed chrome; a titled
+        # frame also reserves a blank line below its title, and buttons
+        # take their own row above the status.
+        chrome = 6 + (1 if self.title else 0) + (1 if self.buttons else 0)
         dialog_h = min(max(self.MIN_HEIGHT, len(flat) + chrome), height)
         visible = max(1, dialog_h - chrome)
         self.page_size = max(1, visible)
@@ -447,7 +450,7 @@ class Frame:
         inner_w = dialog_w - 2
         for line in range(self.scroll, min(len(flat), self.scroll + visible)):
             logical, row, piece = flat[line]
-            y = y0 + 2 + (line - self.scroll)
+            y = y0 + (3 if self.title else 2) + (line - self.scroll)
             selected = logical == self.cursor and row["selectable"]
             if selected:
                 _addstr(scr, y, inner_x, " " * inner_w, theme["bar"])
diff --git a/audiobook.py b/audiobook.py
index 9bdad19..b4cd89d 100755
--- a/audiobook.py
+++ b/audiobook.py
@@ -59,6 +59,7 @@ def convert(backend: str = None, voice: str = None, clone: str = None,
             model_id: str = None, instructions: str = None,
             request_options: dict = None, input_dir: Path = None,
             output_dir: Path = None, api_url: str = None,
+            speaker: str = None,
             progress=None, cancel=None, confirm=None,
             book_files=None, planned=None) -> int:
     """Run one conversion pass with explicit options (used by the CLI and hub).
@@ -101,7 +102,7 @@ def convert(backend: str = None, voice: str = None, clone: str = None,
         book_files, planned = AudiobookConverter.preflight_overwrites(
             backend=backend, voice=voice, voice_mode=voice_mode,
             voice_clone_ref_audio=clone, output_format=output_format,
-            instructions=instructions, confirm=confirm,
+            instructions=instructions, speaker=speaker, confirm=confirm,
         )
     if not book_files:
         print("[INFO] Nothing to convert. Add a .txt, .pdf, or .epub file "
@@ -120,6 +121,7 @@ def convert(backend: str = None, voice: str = None, clone: str = None,
             language=language, backend=backend, voice=voice, debug=debug,
             model_id=model_id, instructions=instructions,
             request_options=request_options, api_url=api_url,
+            speaker=speaker,
             progress=progress, cancel=cancel,
         )
         converter._book_files = book_files
@@ -247,11 +249,21 @@ Examples:
               "a key in the server's voices.json ('default' when it was started "
               "with --ref-audio). audiocpp: a voice_preset or voice_dir entry "
               "(cloning); required for audio.cpp families without built-in "
-              "speakers (everything except Qwen3-TTS CustomVoice). Not used by "
+              "speakers (everything except Qwen3-TTS CustomVoice). When "
+              "AUDIOCPP_CLONE_MODEL_ID names a second server entry (typically "
+              "the Qwen Base model), --voice reroutes to it. Not used by "
               "the qwen backend (use app/converter/config.py SPEAKER or --clone "
               "there).")
     )
     parser.add_argument(
+        "--speaker", type=str, default=None, metavar="NAME",
+        help=("audiocpp only: a built-in Qwen3-TTS CustomVoice speaker name "
+              "(e.g. Vivian, Ryan, Uncle Fu). Selects speaker mode on the "
+              "CustomVoice model entry; mutually exclusive with --voice. "
+              "Without --voice/--speaker, a CustomVoice entry defaults to "
+              "the SPEAKER in app/converter/config.py.")
+    )
+    parser.add_argument(
         "--debug", action="store_true",
         help=("Troubleshooting mode: dump each chunk's raw audio and the exact text "
               "sent for it under the debug/ folder (organized per book and chapter), "
@@ -326,6 +338,11 @@ Examples:
                   "uses a voice configured on the server (voice_presets or "
                   "voice_dir in its config); select it with --voice (see README)")
             args.clone = None
+        if args.voice and args.speaker:
+            parser.error("--voice and --speaker are mutually exclusive with "
+                         "--backend audiocpp: --voice selects a server-side "
+                         "preset (cloning), --speaker a built-in CustomVoice "
+                         "speaker (see README)")
         if args.transcription or args.no_transcription:
             print("[WARNING] --transcription/--no-transcription are ignored with "
                   "--backend audiocpp: the reference transcript is configured on "
@@ -356,6 +373,10 @@ Examples:
         parser.error("--model requires --backend audiocpp; it selects an "
                      "audio.cpp server model entry id")
 
+    if args.speaker is not None and args.backend != BACKEND_AUDIOCPP:
+        parser.error("--speaker requires --backend audiocpp; it selects a "
+                     "built-in Qwen3-TTS CustomVoice speaker")
+
     if args.instructions is not None and args.backend != BACKEND_AUDIOCPP:
         parser.error("--instructions requires --backend audiocpp; it is "
                      "sent as the audio.cpp request's instructions field")
@@ -390,6 +411,7 @@ Examples:
         request_options=request_options,
         input_dir=args.input, output_dir=args.output,
         api_url=api_url,
+        speaker=args.speaker,
     ))