diff options
Diffstat (limited to 'app/converter')
| -rw-r--r-- | app/converter/audio.py | 47 | ||||
| -rw-r--r-- | app/converter/clients/__init__.py | 68 | ||||
| -rw-r--r-- | app/converter/clients/audiocpp.py (renamed from app/converter/tts.py) | 783 | ||||
| -rw-r--r-- | app/converter/clients/base.py | 155 | ||||
| -rw-r--r-- | app/converter/clients/faster.py | 123 | ||||
| -rw-r--r-- | app/converter/clients/languages.py | 74 | ||||
| -rw-r--r-- | app/converter/clients/qwen.py | 322 | ||||
| -rw-r--r-- | app/converter/clients/speakers.py | 57 | ||||
| -rw-r--r-- | app/converter/clients/transcribe.py | 54 | ||||
| -rw-r--r-- | app/converter/converter.py | 21 |
10 files changed, 909 insertions, 795 deletions
diff --git a/app/converter/audio.py b/app/converter/audio.py index eb970ff..90ce5ca 100644 --- a/app/converter/audio.py +++ b/app/converter/audio.py @@ -1,4 +1,9 @@ -"""Audio assembly: combining chunks, speed adjustment, cleanup.""" +"""Audio assembly: combining chunks, speed adjustment, cleanup. + +Every function that touches the run's scratch audio takes its folder as an +explicit CHUNKS_DIR argument — the converter owns the folder constants and +threads them through, so there is no module-global path to mutate. +""" import logging import re @@ -13,8 +18,6 @@ from . import config logger = logging.getLogger(__name__) -CHUNKS_FOLDER = Path(__file__).resolve().parent.parent.parent / "app" / "chunks" - # Tolerance for "is this speed 1.0?" comparisons (banner display, atempo # filter elision); shared by every speed check. SPEED_EPSILON = 1e-6 @@ -362,6 +365,7 @@ def _collect_chunk_files(total_chunks: int, def combine_chunks(total_chunks: int, output_path: Path, chunk_results: Dict[int, Optional[Path]], + *, chunks_dir: Path, speed: float = 1.0, output_format: str = config.AUDIO_FORMAT, intermediate: bool = False, meta: Optional[TrackMeta] = None, @@ -369,14 +373,15 @@ def combine_chunks(total_chunks: int, output_path: Path, """Combine audio chunks into the final audiobook using ffmpeg's concat demuxer. ``chunk_results`` maps chunk numbers to the audio file each chunk produced - (None for failed chunks); failed and missing chunks are skipped. When - ``speed`` differs from 1.0, an additional speed-adjusted copy is written - next to the normal-speed file. ``meta``/``cover`` embed tags and cover - art into the output (skipped for intermediate chapter scratch audio). - Chunks are streamed by ffmpeg, so the whole book is never held in - memory. Set ``intermediate`` for scratch chapter audio on the way to a - larger output (e.g. a chaptered m4b) so save messages don't present it - as the final audiobook. + (None for failed chunks); failed and missing chunks are skipped. The + concat scratch list is written to ``chunks_dir``. When ``speed`` differs + from 1.0, an additional speed-adjusted copy is written next to the + normal-speed file. ``meta``/``cover`` embed tags and cover art into the + output (skipped for intermediate chapter scratch audio). Chunks are + streamed by ffmpeg, so the whole book is never held in memory. Set + ``intermediate`` for scratch chapter audio on the way to a larger output + (e.g. a chaptered m4b) so save messages don't present it as the final + audiobook. """ if shutil.which("ffmpeg") is None or shutil.which("ffprobe") is None: logger.error("ffmpeg and ffprobe are required to combine audio chunks (install ffmpeg)") @@ -391,7 +396,7 @@ def combine_chunks(total_chunks: int, output_path: Path, if missing_chunks: logger.warning("Missing chunks: %s", missing_chunks) - concat_list = CHUNKS_FOLDER / "_concat_list.txt" + concat_list = Path(chunks_dir) / "_concat_list.txt" try: with open(concat_list, "w", encoding="utf-8") as list_file: for chunk_file in chunk_files: @@ -463,12 +468,12 @@ def combine_chunks(total_chunks: int, output_path: Path, pass -def cleanup_chunks() -> None: - """Remove temporary chunk and chapter files from the scratch folder.""" +def cleanup_chunks(chunks_dir: Path) -> None: + """Remove temporary chunk and chapter files from the CHUNKS_DIR scratch folder.""" try: chunk_count = 0 for pattern in ("chunk_*", "chapter_*"): - for chunk_file in CHUNKS_FOLDER.glob(pattern): + for chunk_file in Path(chunks_dir).glob(pattern): try: if chunk_file.is_file(): chunk_file.unlink() @@ -532,13 +537,15 @@ def build_ffmetadata(chapters: List[tuple], path: Path) -> None: def combine_chapters_to_m4b(chapter_files: List[Path], titles: List[str], - output_path: Path, speed: float = 1.0, + output_path: Path, *, chunks_dir: Path, + speed: float = 1.0, meta: Optional[TrackMeta] = None, cover: Optional[Path] = None) -> bool: """Concatenate per-chapter audio into a single m4b with embedded chapter markers. Chapter start/end times are derived from each chapter file's duration and - written as ffmpeg chapter metadata. ``meta``/``cover`` embed tags and + written as ffmpeg chapter metadata; the concat list and metadata scratch + files are written to ``chunks_dir``. ``meta``/``cover`` embed tags and cover art. When ``speed`` differs from 1.0, a speed-adjusted copy (with rescaled chapter markers) is written alongside the normal-speed file. """ @@ -550,9 +557,9 @@ def combine_chapters_to_m4b(chapter_files: List[Path], titles: List[str], logger.error("No chapter files provided") return False - concat_list = CHUNKS_FOLDER / "_concat_list.txt" - metadata_file = CHUNKS_FOLDER / "_chapters.txt" - speed_metadata_file = CHUNKS_FOLDER / "_chapters_speed.txt" + concat_list = Path(chunks_dir) / "_concat_list.txt" + metadata_file = Path(chunks_dir) / "_chapters.txt" + speed_metadata_file = Path(chunks_dir) / "_chapters_speed.txt" try: chapters = [] start_ms = 0 diff --git a/app/converter/clients/__init__.py b/app/converter/clients/__init__.py new file mode 100644 index 0000000..e216011 --- /dev/null +++ b/app/converter/clients/__init__.py @@ -0,0 +1,68 @@ +"""TTS client implementations — one module per backend server. + +Public API: the three client classes (QwenTTSClient, FasterTTSClient, +AudioCppTTSClient), the backend/voice-mode vocabulary, and the shared +helpers (normalize_language, speaker tables, whisper transcription) that +the UIs and setup wizards build on. +""" + +# The TTS backends a conversion can use, in Convert-form order. Each has a +# client module here; the backends package mirrors these keys for its +# install/setup wizards. +BACKEND_QWEN = "qwen" +BACKEND_FASTER = "faster" +BACKEND_AUDIOCPP = "audiocpp" +BACKENDS = (BACKEND_AUDIOCPP, BACKEND_QWEN, BACKEND_FASTER) + +from .base import BaseTTSClient, ConversionCancelled, VOICE_MODE_CLONE, \ + VOICE_MODE_CUSTOM, VOICE_MODES, resolve_request_seed +from .languages import LANGUAGE_ISO_CODES, TTS_LANGUAGES, \ + TTS_LANGUAGE_ALIASES, normalize_language +from .speakers import QWEN3_TTS_SPEAKERS, SPEAKER_DISPLAY_NAMES, \ + is_builtin_speaker, speaker_display_name, speaker_display_name_for +from .transcribe import transcribe_reference_audio, whisper_backend_available +from .qwen import CUSTOM_VOICE_MODEL_ID, MODEL_SIZE, QwenTTSClient +from .faster import SAMPLE_RATE, FasterTTSClient +from .audiocpp import ( + AUDIOCPP_DEFAULT_FAMILY_PROFILE, + AUDIOCPP_FAMILY_PROFILES, + AUDIOCPP_FAMILY_QWEN3_TTS, + AUDIOCPP_LANG_DISPLAY, + AUDIOCPP_LANG_ISO, + AUDIOCPP_LANG_OMIT, + AUDIOCPP_SYNTHESIS_TASKS, + AUDIOCPP_TASK_TTS, + AUDIOCPP_TASK_VDES, + AUDIOCPP_VOICE_CLONE, + AUDIOCPP_VOICE_DESIGN, + AUDIOCPP_VOICE_SPEAKER, + AudioCppFamilyProfile, + AudioCppTTSClient, + audiocpp_entry_voice_capability, +) + +__all__ = [ + # vocabulary + "BACKEND_QWEN", "BACKEND_FASTER", "BACKEND_AUDIOCPP", "BACKENDS", + "VOICE_MODE_CUSTOM", "VOICE_MODE_CLONE", "VOICE_MODES", + # clients + "BaseTTSClient", "ConversionCancelled", "resolve_request_seed", + "QwenTTSClient", "FasterTTSClient", "AudioCppTTSClient", + # model facts + "MODEL_SIZE", "CUSTOM_VOICE_MODEL_ID", "SAMPLE_RATE", + # languages + "TTS_LANGUAGES", "TTS_LANGUAGE_ALIASES", "LANGUAGE_ISO_CODES", + "normalize_language", + # speakers + "QWEN3_TTS_SPEAKERS", "SPEAKER_DISPLAY_NAMES", + "speaker_display_name", "speaker_display_name_for", "is_builtin_speaker", + # transcription + "transcribe_reference_audio", "whisper_backend_available", + # audio.cpp family profiles + "AUDIOCPP_LANG_DISPLAY", "AUDIOCPP_LANG_ISO", "AUDIOCPP_LANG_OMIT", + "AUDIOCPP_FAMILY_QWEN3_TTS", "AUDIOCPP_TASK_TTS", "AUDIOCPP_TASK_VDES", + "AUDIOCPP_SYNTHESIS_TASKS", "AUDIOCPP_VOICE_SPEAKER", + "AUDIOCPP_VOICE_CLONE", "AUDIOCPP_VOICE_DESIGN", + "AudioCppFamilyProfile", "AUDIOCPP_DEFAULT_FAMILY_PROFILE", + "AUDIOCPP_FAMILY_PROFILES", "audiocpp_entry_voice_capability", +] diff --git a/app/converter/tts.py b/app/converter/clients/audiocpp.py index 8130a44..4a161cb 100644 --- a/app/converter/tts.py +++ b/app/converter/clients/audiocpp.py @@ -1,121 +1,25 @@ -"""Client wrappers for the TTS backends. +"""Client for the audio.cpp audiocpp_server (native ggml TTS families).""" -QwenTTSClient talks to the Qwen3-TTS demo server (custom voice / voice clone). -FasterTTSClient talks to the OpenAI-compatible server from the -faster-qwen3-tts repository (voice cloning only; the reference voice is -configured server-side — see the "Faster backend" section of the README). -AudioCppTTSClient talks to the audiocpp_server from the audio.cpp -repository, which can host any TTS model family audio.cpp supports -(Qwen3-TTS, Higgs Audio, VoxCPM2, IndexTTS2, ...) through one OpenAI-style -API; the family is detected from the server at startup (see the -"audio.cpp backend" sections of the README). -""" - -import contextlib -import io import json import logging -import random import shutil -import sys import tempfile -import threading -import time import urllib.error import urllib.parse import urllib.request -import wave from pathlib import Path -from typing import Any, Dict, List, Optional, Tuple +from typing import Any, Dict, List, Optional -from . import config -from .audio import concat_audio_files -from .chunking import split_into_chunks +from .. import config +from ..audio import concat_audio_files +from ..chunking import split_into_chunks +from .base import BaseTTSClient, ConversionCancelled, resolve_request_seed +from .languages import LANGUAGE_ISO_CODES, normalize_language +from .speakers import (is_builtin_speaker, speaker_display_name, + speaker_display_name_for) logger = logging.getLogger(__name__) - -class ConversionCancelled(Exception): - """Raised inside a conversion whose cancel event was set. - - The TUI run view sets a ``threading.Event`` on the TTS client (and the - converter checks it between chunks/chapters/books); the retry loops - raise this so the cancellation propagates out of a sleeping or retrying - request promptly instead of finishing the retry ladder. - """ - - -# Voice modes (re-exported for the CLI and the converter orchestrator). -VOICE_MODE_CUSTOM = "custom_voice" -VOICE_MODE_CLONE = "voice_clone" -VOICE_MODES = (VOICE_MODE_CUSTOM, VOICE_MODE_CLONE) - -# TTS backends (re-exported for the CLI and the converter orchestrator). -BACKEND_QWEN = "qwen" -BACKEND_FASTER = "faster" -BACKEND_AUDIOCPP = "audiocpp" -BACKENDS = (BACKEND_AUDIOCPP, BACKEND_QWEN, BACKEND_FASTER) - -# Languages understood by the Qwen3-TTS API. Display names must match the -# demo dropdown exactly (the demo silently falls back to "Auto" for -# unrecognized values, so languages are validated client-side first). -TTS_LANGUAGES = ( - "Auto", - "Chinese", - "English", - "German", - "Italian", - "Portuguese", - "Spanish", - "Japanese", - "Korean", - "French", - "Russian", -) - -# Short aliases accepted on the command line (ISO 639-1 codes and common -# shorthands), mapped to the display names above. -TTS_LANGUAGE_ALIASES = { - "zh": "Chinese", - "en": "English", - "de": "German", - "it": "Italian", - "pt": "Portuguese", - "es": "Spanish", - "ja": "Japanese", - "ko": "Korean", - "fr": "French", - "ru": "Russian", - "zh-cn": "Chinese", - "zh-tw": "Chinese", - "pt-br": "Portuguese", - "en-us": "English", - "en-gb": "English", -} - -# Qwen display names -> ISO 639-1 codes, for audio.cpp families whose -# language request option takes a code instead of a display name. "Auto" -# has no code and maps to None so the field is omitted and the server -# applies its own default. -LANGUAGE_ISO_CODES = { - "Chinese": "zh", - "English": "en", - "German": "de", - "Italian": "it", - "Portuguese": "pt", - "Spanish": "es", - "Japanese": "ja", - "Korean": "ko", - "French": "fr", - "Russian": "ru", -} - -# --- audio.cpp model families --------------------------------------------- -# -# audiocpp_server exposes the same OpenAI-style API for every TTS family it -# hosts; families only differ in a few request conventions, captured here as -# profiles. Families that are not listed use the default profile below. - # How the "language" request field is expressed by a family. AUDIOCPP_LANG_DISPLAY = "display" # Qwen display names, e.g. "English" AUDIOCPP_LANG_ISO = "iso" # ISO 639-1 codes, e.g. "en" @@ -210,664 +114,8 @@ def audiocpp_entry_voice_capability(family: str, task: str, 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", - "serena": "Serena", - "vivian": "Vivian", - "uncle_fu": "Uncle Fu", - "aiden": "Aiden", - "ono_anna": "Ono Anna", - "sohee": "Sohee", - "eric": "Eric", - "dylan": "Dylan", -} - -# Fixed model facts: both demos run the 1.7B model (the CustomVoice demo -# takes its full HuggingFace id), and the 12Hz codec outputs 24 kHz audio. -MODEL_SIZE = "1.7B" -CUSTOM_VOICE_MODEL_ID = "Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice" -SAMPLE_RATE = 24000 - -CHUNKS_FOLDER = Path(__file__).resolve().parent.parent.parent / "app" / "chunks" - - -def _resolve_request_seed() -> int: - """Resolve the seed sent with every request. - - Returns config.SEED as-is, or (with CONSTANT_SEED and SEED < 0) one - random value drawn per run, meant to be reused for every request so - the voice stays consistent across chunk boundaries. Without - CONSTANT_SEED, -1 is returned so the server re-samples the voice on - every generation. - """ - seed = config.SEED - if config.CONSTANT_SEED and seed < 0: - seed = random.randrange(2 ** 31) - 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 --voice / - Speaker-picker value into what audiocpp_server expects in the request's - voice field. - """ - return SPEAKER_DISPLAY_NAMES.get((name or "").lower(), name) - - -def is_builtin_speaker(name: Optional[str]) -> bool: - """True when NAME is one of the Qwen3-TTS CustomVoice built-in speakers. - - Matches case-insensitively across the canonical ("Uncle_Fu"), display - ("Uncle Fu") and shorthand ("uncle_fu") forms, so the --voice flag and - the Convert form's Speaker picker resolve to the same set. - """ - if not name: - return False - norm = name.lower().replace("_", " ").replace("-", " ") - return any(norm == speaker.lower().replace("_", " ") - for speaker in QWEN3_TTS_SPEAKERS) - - -def speaker_display_name() -> str: - """Return the display name for the configured custom speaker.""" - return speaker_display_name_for(config.SPEAKER) - - -def normalize_language(value: Optional[str]) -> str: - """Normalize a user-provided language name to a Qwen3-TTS display name. - - Accepts the display names in TTS_LANGUAGES case-insensitively as - well as the short aliases in TTS_LANGUAGE_ALIASES (ISO 639-1 codes - and common shorthands). Raises ValueError for anything else, since the - Qwen3-TTS demo silently falls back to "Auto" for unrecognized languages. - """ - if value is None: - raise ValueError("Language must not be None") - candidate = value.strip() - if not candidate: - raise ValueError("Language must not be empty") - for name in TTS_LANGUAGES: - if candidate.lower() == name.lower(): - return name - alias = TTS_LANGUAGE_ALIASES.get(candidate.lower()) - if alias: - return alias - raise ValueError( - f"Unknown language: {value!r}. Expected one of " - f"{', '.join(TTS_LANGUAGES)} (or an alias: " - f"{', '.join(sorted(TTS_LANGUAGE_ALIASES))})." - ) - - -def transcribe_reference_audio(audio_path: str, model_name: str = "base") -> Optional[str]: - """Transcribe reference audio locally using an optional Whisper backend. - - The current qwen-tts demo does not expose a transcription endpoint, so - transcription is done client-side when a Whisper package is available. - Returns None if no backend is installed. - """ - for backend in ("faster_whisper", "whisper"): - try: - if backend == "faster_whisper": - from faster_whisper import WhisperModel - model = WhisperModel(model_name, device="cpu", compute_type="int8") - segments, _ = model.transcribe(audio_path) - text = " ".join(seg.text.strip() for seg in segments).strip() - else: - import whisper - model = whisper.load_model(model_name) - result = model.transcribe(audio_path) - text = (result.get("text") or "").strip() - if text: - logger.info("Transcription complete via %s: %s", backend, text) - return text - except ImportError: - continue - except Exception as exc: - logger.warning("%s transcription failed: %s", backend, exc) - logger.warning("No Whisper backend available; transcription skipped.") - return None - - -def whisper_backend_available() -> Optional[str]: - """Return the name of an importable Whisper backend, or None. - - Checks faster_whisper first (preferred), then the openai-whisper - package, without importing the heavy model code: a bare import probe - is enough to tell whether the package is installed in the current - environment. Used by the make_audiocpp_server_json tool to warn when - neither is present (e.g. the wrong conda environment is active). - """ - for backend in ("faster_whisper", "whisper"): - try: - __import__(backend) - except ImportError: - continue - return backend - return None - - -class _BaseTTSClient: - """Shared chunk retry logic, heartbeat, and chunk file bookkeeping.""" - - # Set by the converter when the run is cancellable (the TUI run view): - # a threading.Event that, once set, aborts the run between requests - # (and interrupts retry back-off sleeps). ``quiet`` silences console - # prints (the run view owns the screen). - cancel = None - quiet = False - - def _report(self, message: str) -> None: - """Print a console line unless quiet (the run view owns the screen).""" - if not self.quiet: - print(message) - - def generate_chunk(self, text: str, chunk_num: int) -> Optional[str]: - """Generate one audio chunk; returns its path in the chunks folder.""" - raise NotImplementedError - - def _cancel_requested(self) -> bool: - """True when the run's cancel event has been set (if any).""" - return isinstance(self.cancel, threading.Event) \ - and self.cancel.is_set() - - def _check_cancelled(self) -> None: - """Raise ConversionCancelled when the cancel event is set.""" - if self._cancel_requested(): - raise ConversionCancelled("Cancelled by user") - - def _sleep(self, seconds: float) -> None: - """Sleep SECONDS, cut short (raising) when the cancel event sets.""" - if isinstance(self.cancel, threading.Event): - if self.cancel.wait(seconds): - raise ConversionCancelled("Cancelled by user") - else: - time.sleep(seconds) - - def _chunk_path(self, chunk_num: int, suffix: str) -> Path: - """Resolve the target path for a chunk, removing stale files first. - - Any stale chunk file for this index is removed so a retry or extension - change can never leave two files matching chunk_NNNN.*. - """ - for stale in CHUNKS_FOLDER.glob(f"chunk_{chunk_num:04d}.*"): - try: - stale.unlink() - except OSError as exc: - logger.debug("Could not remove stale chunk file %s: %s", stale, exc) - return CHUNKS_FOLDER / f"chunk_{chunk_num:04d}{suffix}" - - def process_chunk_with_retry(self, chunk_num: int, text: str) -> Optional[Path]: - """Process a chunk with retry logic. - - Returns the generated chunk file's path, or None when all attempts - failed. Raises ConversionCancelled when the run was cancelled. - """ - for attempt in range(config.MAX_RETRIES): - self._check_cancelled() - try: - result = self.generate_chunk(text, chunk_num) - if result and Path(result).exists(): - return Path(result) - logger.warning("Chunk %d attempt %d failed", chunk_num, attempt + 1) - except ConversionCancelled: - raise - except Exception as exc: - logger.warning("Chunk %d attempt %d error: %s", chunk_num, attempt + 1, exc) - - if attempt < config.MAX_RETRIES - 1: - sleep_time = 5 + (2 ** attempt) - logger.info("Waiting %ds before retry...", sleep_time) - self._sleep(sleep_time) - - logger.error("Chunk %d failed after %d attempts", chunk_num, config.MAX_RETRIES) - return None - - @contextlib.contextmanager - def _chunk_heartbeat(self, chunk_num: int): - """Log a periodic "still working" record while a request generates.""" - stop = threading.Event() - subject = f"Chunk {chunk_num}" - - def _beat(): - start = time.time() - while not stop.wait(config.HEARTBEAT_INTERVAL_SECONDS): - elapsed = time.time() - start - if self.quiet: - logger.info("%s still generating — %dm %ds elapsed", - subject, int(elapsed // 60), int(elapsed % 60)) - else: - print(f"[...] {subject} still generating — " - f"{int(elapsed // 60)}m {int(elapsed % 60)}s elapsed", - flush=True) - - thread = threading.Thread(target=_beat, daemon=True) - thread.start() - try: - yield - finally: - stop.set() - thread.join() - - -class QwenTTSClient(_BaseTTSClient): - """Generates audio chunks through a Qwen3-TTS demo server.""" - - def __init__(self, voice_mode: str = "custom_voice", voice_clone_ref_audio: Optional[str] = None, - voice_clone_ref_text: Optional[str] = None, skip_transcription: bool = False, - language: Optional[str] = None, api_url: Optional[str] = None, - quiet: bool = False): - # Quiet before connecting so connect-time status lines never reach - # a screen the TUI run view owns. - self.quiet = bool(quiet) - if voice_mode not in VOICE_MODES: - raise ValueError( - f"Unknown voice mode: {voice_mode!r} (expected one of {VOICE_MODES})" - ) - self.voice_mode = voice_mode - self.voice_clone_ref_audio = voice_clone_ref_audio - self.voice_clone_ref_text = (voice_clone_ref_text or "").strip() - self.skip_transcription = skip_transcription - # api_url overrides the configured endpoint for the active voice mode - # (used by the hub's "[remote]" backend entries and --api-url). - self.api_url = (api_url or "").strip() or None - # Seed sent with every request: config.SEED as-is, or (with - # CONSTANT_SEED and SEED < 0) one random value drawn per run and - # reused for every request so the voice stays consistent across - # chunk boundaries. Without CONSTANT_SEED, -1 is forwarded so the - # server re-samples the voice on every generation. - self._seed = _resolve_request_seed() - if language is None: - language = config.LANGUAGE - # Validate before connecting so bad values fail fast without a server. - self.language = normalize_language(language) - self.client = None - self.api_info: Dict[str, Any] = {} - self.clone_client = None - self.clone_api_info: Dict[str, Any] = {} - self._ref_audio_filedata: Optional[Dict[str, Any]] = None - self._connect() - - # ------------------------------------------------------------------ - # Connection - # ------------------------------------------------------------------ - - def _connect(self) -> None: - api_url = self.api_url or ( - config.CLONE_API_URL if self.voice_mode == VOICE_MODE_CLONE - else config.QWEN_API_URL) - try: - if self.voice_mode == VOICE_MODE_CLONE: - # Voice clone uses the Base-model demo, which is a separate server - # from the CustomVoice demo (that one only exposes /run_instruct). - self._init_client(api_url, clone=True) - self._report(f"[OK] Connected to Voice Clone API at {api_url}") - self._resolve_reference_text() - else: - self._init_client(api_url, clone=False) - self._report("[OK] Connected to Qwen API") - except Exception as exc: - raise RuntimeError( - f"Qwen API initialization failed at {api_url}: {exc}. " - "Make sure the Qwen demo server is running and reachable, and that your " - "installed Qwen3-TTS version matches this converter's API expectations " - "(voice clone requires the Base-model demo: Qwen/Qwen3-TTS-12Hz-1.7B-Base)." - ) from exc - - def _resolve_reference_text(self) -> None: - """Resolve the reference transcript: explicit text, then local - transcription, then x-vector-only mode.""" - if not self.voice_clone_ref_text and self.voice_clone_ref_audio: - if self.skip_transcription: - self._report("[INFO] Skipping reference audio transcription (--no-transcription).") - else: - self._report("[INFO] Transcribing reference audio for voice cloning...") - self.voice_clone_ref_text = self.transcribe_audio(self.voice_clone_ref_audio) or "" - if not self.voice_clone_ref_text: - self._report("[WARNING] No reference text available; using " - "x-vector-only clone mode (lower quality).") - self._report(' Pass --transcription "..." for higher-quality in-context cloning.') - else: - self._report(f"[OK] Reference text:\n{self.voice_clone_ref_text}") - - def _init_client(self, url: str, clone: bool = False) -> None: - """Initialize a Gradio client and store its API metadata. - - gradio_client prints its usage info directly to stdout while the - client is created and its API metadata loaded, so stdout is swapped - for a buffer for the whole process; the captured text is re-emitted - at DEBUG level for troubleshooting. - """ - from gradio_client import Client - - logger.info("Connecting to Qwen API at %s...", url) - old_stdout = sys.stdout - captured = io.StringIO() - sys.stdout = captured - try: - try: - client = Client(url, httpx_kwargs={"timeout": config.API_TIMEOUT}) - except TypeError: - # Older gradio_client versions don't support httpx_kwargs. - client = Client(url) - if clone: - self.clone_client = client - self.clone_api_info = self._load_api_info(client) - else: - self.client = client - self.api_info = self._load_api_info(client) - finally: - sys.stdout = old_stdout - usage_info = captured.getvalue().strip() - if usage_info: - logger.debug("Gradio client output for %s:\n%s", url, usage_info) - logger.info("Connected to Qwen API") - - @staticmethod - def _load_api_info(client) -> Dict[str, Any]: - """Load available API metadata from the Gradio app.""" - try: - return client.view_api(return_format="dict") - except Exception as exc: - logger.warning("Unable to read API metadata: %s", exc) - return {} - - def _resolve_api_name(self, *candidates: str, api_info: Optional[Dict[str, Any]] = None) -> str: - """Return the first available api_name from candidate list.""" - info = api_info if api_info is not None else self.api_info - named_endpoints = info.get("named_endpoints", {}) - for candidate in candidates: - if candidate in named_endpoints: - return candidate - return candidates[0] - - def _endpoint_accepts_param(self, api_name: str, param_name: str, - api_info: Optional[Dict[str, Any]] = None) -> bool: - """Check whether endpoint input schema includes the given parameter.""" - info = api_info if api_info is not None else self.api_info - endpoint = info.get("named_endpoints", {}).get(api_name, {}) - parameters = endpoint.get("parameters", []) - return any(parameter.get("parameter_name") == param_name for parameter in parameters) - - # ------------------------------------------------------------------ - # Reference audio transcription (voice clone) - # ------------------------------------------------------------------ - - def transcribe_audio(self, audio_path: str) -> Optional[str]: - """Transcribe reference audio locally using an optional Whisper backend.""" - return transcribe_reference_audio(audio_path) - - # ------------------------------------------------------------------ - # Chunk generation - # ------------------------------------------------------------------ - - def generate_chunk(self, text: str, chunk_num: int) -> Optional[str]: - """Generate one audio chunk; returns its path in the chunks folder. - - The text is split into sub-requests of at most - ``config.CHUNK_SIZE`` words each (the book-level chunker - normally guarantees this already; the split is defense in depth - against pathological input such as a punctuation-free run of - text), and the audio files returned for the sub-requests are - concatenated into one chunk file. - """ - try: - sub_texts = split_into_chunks(text, max_words=config.CHUNK_SIZE) - if not sub_texts: - raise RuntimeError("No text to synthesize") - - output_path: Optional[Path] = None - with tempfile.TemporaryDirectory(prefix="tts_parts_") as parts_dir, \ - self._chunk_heartbeat(chunk_num): - part_paths = [ - self._generate_sub_request(sub_text, parts_dir, sub_num, - len(sub_texts), chunk_num) - for sub_num, sub_text in enumerate(sub_texts, 1) - ] - if len(part_paths) == 1: - suffix = part_paths[0].suffix or ".wav" - output_path = self._chunk_path(chunk_num, suffix) - shutil.copy2(part_paths[0], output_path) - else: - output_path = self._chunk_path(chunk_num, ".wav") - concat_audio_files(part_paths, output_path) - - logger.debug("Chunk %d generated successfully (%d sub-request(s))", - chunk_num, len(sub_texts)) - return str(output_path) - - except ConversionCancelled: - raise - except Exception as exc: - logger.error("Qwen chunk processing failed for chunk %d: %s", chunk_num, exc) - return None - - def _generate_sub_request(self, text: str, parts_dir: str, sub_num: int, - sub_total: int, chunk_num: int) -> Path: - """Run one API generation for ``text``; returns the downloaded audio.""" - if sub_total > 1: - logger.info("Chunk %d: oversized input split into %d requests " - "(sub-request %d/%d)", chunk_num, sub_total, sub_num, sub_total) - if self.voice_mode == VOICE_MODE_CUSTOM: - result = self._generate_custom_voice(text) - elif self.voice_mode == VOICE_MODE_CLONE: - result = self._generate_voice_clone(text) - else: - raise ValueError(f"Unknown voice mode: {self.voice_mode}") - - if not isinstance(result, (tuple, list)) or not result: - raise RuntimeError("Qwen API returned an invalid result") - - audio_path = result[0] # First element is the audio file path - if not isinstance(audio_path, (str, Path)) or not audio_path: - raise RuntimeError("Qwen API did not return an audio file path") - - source = Path(audio_path) - if not source.exists(): - raise RuntimeError(f"Generated audio file not found: {audio_path}") - - destination = Path(parts_dir) / f"part_{sub_num:02d}{source.suffix or '.wav'}" - shutil.copy2(source, destination) - - return destination - - # ------------------------------------------------------------------ - # API payloads - # ------------------------------------------------------------------ - - def _generate_custom_voice(self, text: str) -> Tuple: - """Generate audio using CustomVoice mode.""" - custom_api = self._resolve_api_name("/run_instruct", "/run_custom_voice", "/generate_custom_voice") - if custom_api == "/run_instruct": - payload = dict( - text=text, - lang_disp=self.language, - spk_disp=speaker_display_name(), - instruct=config.INSTRUCT, - ) - else: - payload = dict( - text=text, - language=self.language, - speaker=config.SPEAKER, - instruct=config.INSTRUCT, - ) - if self._endpoint_accepts_param(custom_api, "model_id_cv"): - payload["model_id_cv"] = CUSTOM_VOICE_MODEL_ID - elif self._endpoint_accepts_param(custom_api, "model_size"): - payload["model_size"] = MODEL_SIZE - - if self._endpoint_accepts_param(custom_api, "seed"): - payload["seed"] = self._seed - - return self.client.predict(**payload, api_name=custom_api) - - def _ref_audio_payload(self) -> Dict[str, Any]: - """Gradio file payload for the reference audio (built once, reused).""" - if self._ref_audio_filedata is None: - from gradio_client import handle_file - self._ref_audio_filedata = handle_file(self.voice_clone_ref_audio) - return self._ref_audio_filedata - - def _generate_voice_clone(self, text: str) -> Tuple: - """Generate audio using Voice Clone mode.""" - if not Path(self.voice_clone_ref_audio).exists(): - raise FileNotFoundError(f"Reference audio not found: {self.voice_clone_ref_audio}") - - if self.clone_client is None: - raise RuntimeError("Voice Clone client is not initialized. Is the Base-model demo running?") - - clone_api = self._resolve_api_name("/run_voice_clone", "/generate_voice_clone", - api_info=self.clone_api_info) - use_xvector = config.XVECTOR_ONLY or not self.voice_clone_ref_text - - if clone_api == "/run_voice_clone": - payload = dict( - ref_aud=self._ref_audio_payload(), - ref_txt=self.voice_clone_ref_text, - use_xvec=use_xvector, - text=text, - lang_disp=self.language, - ) - else: - payload = dict( - ref_audio=self._ref_audio_payload(), - ref_text=self.voice_clone_ref_text, - target_text=text, - language=self.language, - use_xvector_only=use_xvector, - ) - optional_params = { - "model_size": MODEL_SIZE, - "seed": self._seed, - } - for name, value in optional_params.items(): - if self._endpoint_accepts_param(clone_api, name, api_info=self.clone_api_info): - payload[name] = value - - return self.clone_client.predict(**payload, api_name=clone_api) - - -class FasterTTSClient(_BaseTTSClient): - """Generates audio chunks through a faster-qwen3-tts server. - - Talks to the OpenAI-compatible server shipped in the faster-qwen3-tts - repository (examples/openai_server.py). The reference voice (ref audio, - ref text) and language are configured on the server itself via - --ref-audio/--ref-text or a --voices JSON file; this client only sends - text. Unlike the Qwen demo, the server performs one generation per - request, so long chunks are sub-chunked client-side. - """ - - def __init__(self, voice: Optional[str] = None, api_url: Optional[str] = None, - quiet: bool = False): - # Quiet before connecting so connect-time status lines never reach - # a screen the TUI run view owns. - self.quiet = bool(quiet) - self.voice = voice or config.FASTER_VOICE - self.api_url = (api_url or config.FASTER_API_URL).rstrip("/") - self._check_health() - - def _check_health(self) -> None: - """Verify the server is reachable and its model is loaded.""" - url = f"{self.api_url}/health" - try: - with urllib.request.urlopen(url, timeout=10) as response: - payload = json.loads(response.read().decode("utf-8")) - except Exception as exc: - raise RuntimeError( - f"Faster TTS server not reachable at {url}: {exc}. " - "Start the faster-qwen3-tts OpenAI-compatible server first " - "(see the 'Faster backend' section of the README)." - ) from exc - if not payload.get("model_loaded"): - raise RuntimeError( - "The faster TTS server is running but its model is not loaded yet; " - "wait for model download and startup to finish, then retry." - ) - self._report(f"[OK] Connected to faster TTS API at {self.api_url} (voice '{self.voice}')") - self._report(f"[INFO] The server silently falls back to its first configured voice if " - f"'{self.voice}' is not defined in its voice config (see README).") - - # ------------------------------------------------------------------ - # HTTP requests - # ------------------------------------------------------------------ - - def _request_pcm(self, text: str) -> bytes: - """POST one sub-chunk and return raw 16-bit mono PCM bytes.""" - url = f"{self.api_url}/v1/audio/speech" - payload = json.dumps({ - "model": "tts-1", - "input": text, - "voice": self.voice, - "response_format": "pcm", - }).encode("utf-8") - request = urllib.request.Request( - url, data=payload, headers={"Content-Type": "application/json"}, method="POST") - try: - with urllib.request.urlopen(request, timeout=config.API_TIMEOUT) as response: - pcm = response.read() - except urllib.error.HTTPError as exc: - detail = "" - try: - detail = exc.read().decode("utf-8", errors="replace")[:200] - except Exception: - pass - raise RuntimeError(f"Faster TTS server returned HTTP {exc.code}: {detail}") from exc - except urllib.error.URLError as exc: - raise RuntimeError(f"Faster TTS request failed: {exc.reason}") from exc - if not pcm: - raise RuntimeError("Faster TTS server returned empty audio") - return pcm - - # ------------------------------------------------------------------ - # Chunk generation - # ------------------------------------------------------------------ - - def generate_chunk(self, text: str, chunk_num: int) -> Optional[str]: - """Generate one audio chunk; returns its path in the chunks folder.""" - try: - sub_chunks = split_into_chunks(text, max_words=config.CHUNK_SIZE) - if not sub_chunks: - raise RuntimeError("No text to synthesize") - - pcm_parts: List[bytes] = [] - with self._chunk_heartbeat(chunk_num): - for sub_num, sub_text in enumerate(sub_chunks, 1): - pcm = self._request_pcm(sub_text) - pcm_parts.append(pcm) - - output_path = self._chunk_path(chunk_num, ".wav") - with wave.open(str(output_path), "wb") as wav_file: - wav_file.setnchannels(1) - wav_file.setsampwidth(2) - wav_file.setframerate(SAMPLE_RATE) - wav_file.writeframes(b"".join(pcm_parts)) - - logger.debug("Chunk %d generated (%d sub-chunks)", chunk_num, len(sub_chunks)) - return str(output_path) - - except ConversionCancelled: - raise - except Exception as exc: - logger.error("Faster chunk processing failed for chunk %d: %s", chunk_num, exc) - return None - -class AudioCppTTSClient(_BaseTTSClient): +class AudioCppTTSClient(BaseTTSClient): """Generates audio chunks through an audio.cpp audiocpp_server. Talks to the OpenAI-style HTTP API of audiocpp_server, which hosts TTS @@ -928,15 +176,14 @@ class AudioCppTTSClient(_BaseTTSClient): used for the Qwen client. """ - def __init__(self, voice: Optional[str] = None, language: Optional[str] = None, + def __init__(self, chunks_dir: Path, + voice: Optional[str] = None, language: Optional[str] = None, api_url: Optional[str] = None, model_id: Optional[str] = None, instructions: Optional[str] = None, request_options: Optional[Dict[str, str]] = None, quiet: bool = False): - # Quiet before connecting so connect-time status lines never reach - # a screen the TUI run view owns. - self.quiet = bool(quiet) + super().__init__(chunks_dir, quiet=quiet) 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 @@ -948,10 +195,10 @@ class AudioCppTTSClient(_BaseTTSClient): self.language = normalize_language( language if language is not None else config.LANGUAGE) # One seed value per run, reused for every request (see - # _resolve_request_seed). Unlike the Qwen demo, audio.cpp has no + # resolve_request_seed). Unlike the Qwen demo, audio.cpp has no # 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() + self._seed = resolve_request_seed() # Voice selection (the --voice name). preset_mode / speaker_mode are # resolved in _connect: a --voice that names a built-in CustomVoice # speaker on a speaker-capable entry selects speaker mode; every diff --git a/app/converter/clients/base.py b/app/converter/clients/base.py new file mode 100644 index 0000000..a0f28cf --- /dev/null +++ b/app/converter/clients/base.py @@ -0,0 +1,155 @@ +"""Shared TTS client plumbing: cancellation, retries, chunk bookkeeping.""" + +import contextlib +import logging +import random +import threading +import time +from pathlib import Path +from typing import Optional + +from .. import config + +logger = logging.getLogger(__name__) + + +class ConversionCancelled(Exception): + """Raised when the run's cancel event is set (between requests).""" + + +# How a run supplies its voice: a built-in CustomVoice speaker, or by +# cloning a reference audio clip (the faster and audiocpp backends always +# clone server-side; only the Qwen client branches on this at request time). +VOICE_MODE_CUSTOM = "custom_voice" +VOICE_MODE_CLONE = "voice_clone" +VOICE_MODES = (VOICE_MODE_CUSTOM, VOICE_MODE_CLONE) + + +def resolve_request_seed() -> int: + """Resolve the seed sent with every request. + + Returns config.SEED as-is, or (with CONSTANT_SEED and SEED < 0) one + random value drawn per run, meant to be reused for every request so + the voice stays consistent across chunk boundaries. Without + CONSTANT_SEED, -1 is returned so the server re-samples the voice on + every generation. + """ + seed = config.SEED + if config.CONSTANT_SEED and seed < 0: + seed = random.randrange(2 ** 31) + return seed + + +class BaseTTSClient: + """Shared chunk retry logic, heartbeat, and chunk file bookkeeping. + + CHUNKS_DIR is the scratch folder the generated chunk files are written + to — provided by the converter that owns the run's folders, never a + module global, so concurrent runs (and tests) cannot step on each other. + """ + + # Class-level defaults so a partially-constructed instance behaves like + # a plain console run (tests build clients via __new__). + cancel = None + quiet = False + + def __init__(self, chunks_dir: Path, quiet: bool = False): + self.chunks_dir = Path(chunks_dir) + # Quiet silences console prints (the run view owns the screen). + self.quiet = bool(quiet) + # Set by the converter when the run is cancellable (the TUI run + # view): a threading.Event that, once set, aborts the run between + # requests (and interrupts retry back-off sleeps). + self.cancel = None + + def _report(self, message: str) -> None: + """Print a console line unless quiet (the run view owns the screen).""" + if not self.quiet: + print(message) + + def generate_chunk(self, text: str, chunk_num: int) -> Optional[str]: + """Generate one audio chunk; returns its path in the chunks folder.""" + raise NotImplementedError + + def _cancel_requested(self) -> bool: + """True when the run's cancel event has been set (if any).""" + return isinstance(self.cancel, threading.Event) \ + and self.cancel.is_set() + + def _check_cancelled(self) -> None: + """Raise ConversionCancelled when the cancel event is set.""" + if self._cancel_requested(): + raise ConversionCancelled("Cancelled by user") + + def _sleep(self, seconds: float) -> None: + """Sleep SECONDS, cut short (raising) when the cancel event sets.""" + if isinstance(self.cancel, threading.Event): + if self.cancel.wait(seconds): + raise ConversionCancelled("Cancelled by user") + else: + time.sleep(seconds) + + def _chunk_path(self, chunk_num: int, suffix: str) -> Path: + """Resolve the target path for a chunk, removing stale files first. + + Any stale chunk file for this index is removed so a retry or extension + change can never leave two files matching chunk_NNNN.*. + """ + for stale in self.chunks_dir.glob(f"chunk_{chunk_num:04d}.*"): + try: + stale.unlink() + except OSError as exc: + logger.debug("Could not remove stale chunk file %s: %s", stale, exc) + return self.chunks_dir / f"chunk_{chunk_num:04d}{suffix}" + + def process_chunk_with_retry(self, chunk_num: int, text: str) -> Optional[Path]: + """Process a chunk with retry logic. + + Returns the generated chunk file's path, or None when all attempts + failed. Raises ConversionCancelled when the run was cancelled. + """ + for attempt in range(config.MAX_RETRIES): + self._check_cancelled() + try: + result = self.generate_chunk(text, chunk_num) + if result and Path(result).exists(): + return Path(result) + logger.warning("Chunk %d attempt %d failed", chunk_num, attempt + 1) + except ConversionCancelled: + raise + except Exception as exc: + logger.warning("Chunk %d attempt %d error: %s", chunk_num, attempt + 1, exc) + + if attempt < config.MAX_RETRIES - 1: + sleep_time = 5 + (2 ** attempt) + logger.info("Waiting %ds before retry...", sleep_time) + self._sleep(sleep_time) + + logger.error("Chunk %d failed after %d attempts", chunk_num, config.MAX_RETRIES) + return None + + @contextlib.contextmanager + def _chunk_heartbeat(self, chunk_num: int): + """Log a periodic "still working" record while a request generates.""" + stop = threading.Event() + subject = f"Chunk {chunk_num}" + + def _beat(): + start = time.time() + while not stop.wait(config.HEARTBEAT_INTERVAL_SECONDS): + elapsed = time.time() - start + if self.quiet: + logger.info("%s still generating — %dm %ds elapsed", + subject, int(elapsed // 60), int(elapsed % 60)) + else: + print(f"[...] {subject} still generating — " + f"{int(elapsed // 60)}m {int(elapsed % 60)}s elapsed", + flush=True) + + thread = threading.Thread(target=_beat, daemon=True) + thread.start() + try: + yield + finally: + stop.set() + thread.join() diff --git a/app/converter/clients/faster.py b/app/converter/clients/faster.py new file mode 100644 index 0000000..df98546 --- /dev/null +++ b/app/converter/clients/faster.py @@ -0,0 +1,123 @@ +"""Client for the faster-qwen3-tts OpenAI-compatible server.""" + +import json +import logging +import urllib.error +import urllib.request +import wave +from pathlib import Path +from typing import List, Optional + +from .. import config +from ..chunking import split_into_chunks +from .base import BaseTTSClient, ConversionCancelled + +logger = logging.getLogger(__name__) + +# The 12Hz codec the faster server synthesizes with outputs 24 kHz audio. +SAMPLE_RATE = 24000 + + +class FasterTTSClient(BaseTTSClient): + """Generates audio chunks through a faster-qwen3-tts server. + + Talks to the OpenAI-compatible server shipped in the faster-qwen3-tts + repository (examples/openai_server.py). The reference voice (ref audio, + ref text) and language are configured on the server itself via + --ref-audio/--ref-text or a --voices JSON file; this client only sends + text. Unlike the Qwen demo, the server performs one generation per + request, so long chunks are sub-chunked client-side. + """ + + def __init__(self, chunks_dir: Path, + voice: Optional[str] = None, api_url: Optional[str] = None, + quiet: bool = False): + super().__init__(chunks_dir, quiet=quiet) + self.voice = voice or config.FASTER_VOICE + self.api_url = (api_url or config.FASTER_API_URL).rstrip("/") + self._check_health() + + def _check_health(self) -> None: + """Verify the server is reachable and its model is loaded.""" + url = f"{self.api_url}/health" + try: + with urllib.request.urlopen(url, timeout=10) as response: + payload = json.loads(response.read().decode("utf-8")) + except Exception as exc: + raise RuntimeError( + f"Faster TTS server not reachable at {url}: {exc}. " + "Start the faster-qwen3-tts OpenAI-compatible server first " + "(see the 'Faster backend' section of the README)." + ) from exc + if not payload.get("model_loaded"): + raise RuntimeError( + "The faster TTS server is running but its model is not loaded yet; " + "wait for model download and startup to finish, then retry." + ) + self._report(f"[OK] Connected to faster TTS API at {self.api_url} (voice '{self.voice}')") + self._report(f"[INFO] The server silently falls back to its first configured voice if " + f"'{self.voice}' is not defined in its voice config (see README).") + + # ------------------------------------------------------------------ + # HTTP requests + # ------------------------------------------------------------------ + + def _request_pcm(self, text: str) -> bytes: + """POST one sub-chunk and return raw 16-bit mono PCM bytes.""" + url = f"{self.api_url}/v1/audio/speech" + payload = json.dumps({ + "model": "tts-1", + "input": text, + "voice": self.voice, + "response_format": "pcm", + }).encode("utf-8") + request = urllib.request.Request( + url, data=payload, headers={"Content-Type": "application/json"}, method="POST") + try: + with urllib.request.urlopen(request, timeout=config.API_TIMEOUT) as response: + pcm = response.read() + except urllib.error.HTTPError as exc: + detail = "" + try: + detail = exc.read().decode("utf-8", errors="replace")[:200] + except Exception: + pass + raise RuntimeError(f"Faster TTS server returned HTTP {exc.code}: {detail}") from exc + except urllib.error.URLError as exc: + raise RuntimeError(f"Faster TTS request failed: {exc.reason}") from exc + if not pcm: + raise RuntimeError("Faster TTS server returned empty audio") + return pcm + + # ------------------------------------------------------------------ + # Chunk generation + # ------------------------------------------------------------------ + + def generate_chunk(self, text: str, chunk_num: int) -> Optional[str]: + """Generate one audio chunk; returns its path in the chunks folder.""" + try: + sub_chunks = split_into_chunks(text, max_words=config.CHUNK_SIZE) + if not sub_chunks: + raise RuntimeError("No text to synthesize") + + pcm_parts: List[bytes] = [] + with self._chunk_heartbeat(chunk_num): + for sub_num, sub_text in enumerate(sub_chunks, 1): + pcm = self._request_pcm(sub_text) + pcm_parts.append(pcm) + + output_path = self._chunk_path(chunk_num, ".wav") + with wave.open(str(output_path), "wb") as wav_file: + wav_file.setnchannels(1) + wav_file.setsampwidth(2) + wav_file.setframerate(SAMPLE_RATE) + wav_file.writeframes(b"".join(pcm_parts)) + + logger.debug("Chunk %d generated (%d sub-chunks)", chunk_num, len(sub_chunks)) + return str(output_path) + + except ConversionCancelled: + raise + except Exception as exc: + logger.error("Faster chunk processing failed for chunk %d: %s", chunk_num, exc) + return None diff --git a/app/converter/clients/languages.py b/app/converter/clients/languages.py new file mode 100644 index 0000000..079dead --- /dev/null +++ b/app/converter/clients/languages.py @@ -0,0 +1,74 @@ +"""Language tables shared by the TTS clients and their UIs.""" + +from typing import Optional + +# Languages the Qwen3-TTS demo accepts as display names (its API silently +# falls back to "Auto" for anything else, so unknown names are rejected +# before a run starts instead of mispronouncing a whole book). +TTS_LANGUAGES = ( + "Auto", "Chinese", "English", "German", "Italian", "Portuguese", + "Spanish", "Japanese", "Korean", "French", "Russian", +) + +# Short aliases accepted on the command line (ISO 639-1 codes and common +# shorthands), mapped to the display names above. +TTS_LANGUAGE_ALIASES = { + "zh": "Chinese", + "en": "English", + "de": "German", + "it": "Italian", + "pt": "Portuguese", + "es": "Spanish", + "ja": "Japanese", + "ko": "Korean", + "fr": "French", + "ru": "Russian", + "zh-cn": "Chinese", + "zh-tw": "Chinese", + "pt-br": "Portuguese", + "en-us": "English", + "en-gb": "English", +} + +# Qwen display names -> ISO 639-1 codes, for audio.cpp families whose +# language request option takes a code instead of a display name. "Auto" +# has no code and maps to None so the field is omitted and the server +# applies its own default. +LANGUAGE_ISO_CODES = { + "Chinese": "zh", + "English": "en", + "German": "de", + "Italian": "it", + "Portuguese": "pt", + "Spanish": "es", + "Japanese": "ja", + "Korean": "ko", + "French": "fr", + "Russian": "ru", +} + + +def normalize_language(value: Optional[str]) -> str: + """Normalize a user-provided language name to a Qwen3-TTS display name. + + Accepts the display names in TTS_LANGUAGES case-insensitively as + well as the short aliases in TTS_LANGUAGE_ALIASES (ISO 639-1 codes + and common shorthands). Raises ValueError for anything else, since the + Qwen3-TTS demo silently falls back to "Auto" for unrecognized languages. + """ + if value is None: + raise ValueError("Language must not be None") + candidate = value.strip() + if not candidate: + raise ValueError("Language must not be empty") + for name in TTS_LANGUAGES: + if candidate.lower() == name.lower(): + return name + alias = TTS_LANGUAGE_ALIASES.get(candidate.lower()) + if alias: + return alias + raise ValueError( + f"Unknown language: {value!r}. Expected one of " + f"{', '.join(TTS_LANGUAGES)} (or an alias: " + f"{', '.join(sorted(TTS_LANGUAGE_ALIASES))})." + ) diff --git a/app/converter/clients/qwen.py b/app/converter/clients/qwen.py new file mode 100644 index 0000000..354ee04 --- /dev/null +++ b/app/converter/clients/qwen.py @@ -0,0 +1,322 @@ +"""Client for the qwen-tts Gradio demo servers (CustomVoice + Base).""" + +import io +import logging +import shutil +import sys +import tempfile +from pathlib import Path +from typing import Any, Dict, Optional, Tuple + +from .. import config +from ..audio import concat_audio_files +from ..chunking import split_into_chunks +from .base import (BaseTTSClient, ConversionCancelled, resolve_request_seed, + VOICE_MODE_CLONE, VOICE_MODE_CUSTOM, VOICE_MODES) +from .languages import normalize_language +from .speakers import speaker_display_name + +logger = logging.getLogger(__name__) + +# Fixed model facts: both demos run the 1.7B model (the CustomVoice demo +# takes its full HuggingFace id), and the 12Hz codec outputs 24 kHz audio. +MODEL_SIZE = "1.7B" +CUSTOM_VOICE_MODEL_ID = "Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice" + + +class QwenTTSClient(BaseTTSClient): + """Generates audio chunks through a Qwen3-TTS demo server.""" + + def __init__(self, chunks_dir: Path, + voice_mode: str = "custom_voice", voice_clone_ref_audio: Optional[str] = None, + voice_clone_ref_text: Optional[str] = None, skip_transcription: bool = False, + language: Optional[str] = None, api_url: Optional[str] = None, + quiet: bool = False): + super().__init__(chunks_dir, quiet=quiet) + if voice_mode not in VOICE_MODES: + raise ValueError( + f"Unknown voice mode: {voice_mode!r} (expected one of {VOICE_MODES})" + ) + self.voice_mode = voice_mode + self.voice_clone_ref_audio = voice_clone_ref_audio + self.voice_clone_ref_text = (voice_clone_ref_text or "").strip() + self.skip_transcription = skip_transcription + # api_url overrides the configured endpoint for the active voice mode + # (used by the hub's "[remote]" backend entries and --api-url). + self.api_url = (api_url or "").strip() or None + # Seed sent with every request: config.SEED as-is, or (with + # CONSTANT_SEED and SEED < 0) one random value drawn per run and + # reused for every request so the voice stays consistent across + # chunk boundaries. Without CONSTANT_SEED, -1 is forwarded so the + # server re-samples the voice on every generation. + self._seed = resolve_request_seed() + if language is None: + language = config.LANGUAGE + # Validate before connecting so bad values fail fast without a server. + self.language = normalize_language(language) + self.client = None + self.api_info: Dict[str, Any] = {} + self.clone_client = None + self.clone_api_info: Dict[str, Any] = {} + self._ref_audio_filedata: Optional[Dict[str, Any]] = None + self._connect() + + # ------------------------------------------------------------------ + # Connection + # ------------------------------------------------------------------ + + def _connect(self) -> None: + api_url = self.api_url or ( + config.CLONE_API_URL if self.voice_mode == VOICE_MODE_CLONE + else config.QWEN_API_URL) + try: + if self.voice_mode == VOICE_MODE_CLONE: + # Voice clone uses the Base-model demo, which is a separate server + # from the CustomVoice demo (that one only exposes /run_instruct). + self._init_client(api_url, clone=True) + self._report(f"[OK] Connected to Voice Clone API at {api_url}") + self._resolve_reference_text() + else: + self._init_client(api_url, clone=False) + self._report("[OK] Connected to Qwen API") + except Exception as exc: + raise RuntimeError( + f"Qwen API initialization failed at {api_url}: {exc}. " + "Make sure the Qwen demo server is running and reachable, and that your " + "installed Qwen3-TTS version matches this converter's API expectations " + "(voice clone requires the Base-model demo: Qwen/Qwen3-TTS-12Hz-1.7B-Base)." + ) from exc + + def _resolve_reference_text(self) -> None: + """Resolve the reference transcript: explicit text, then local + transcription, then x-vector-only mode.""" + if not self.voice_clone_ref_text and self.voice_clone_ref_audio: + if self.skip_transcription: + self._report("[INFO] Skipping reference audio transcription (--no-transcription).") + else: + self._report("[INFO] Transcribing reference audio for voice cloning...") + self.voice_clone_ref_text = self.transcribe_audio(self.voice_clone_ref_audio) or "" + if not self.voice_clone_ref_text: + self._report("[WARNING] No reference text available; using " + "x-vector-only clone mode (lower quality).") + self._report(' Pass --transcription "..." for higher-quality in-context cloning.') + else: + self._report(f"[OK] Reference text:\n{self.voice_clone_ref_text}") + + def _init_client(self, url: str, clone: bool = False) -> None: + """Initialize a Gradio client and store its API metadata. + + gradio_client prints its usage info directly to stdout while the + client is created and its API metadata loaded, so stdout is swapped + for a buffer for the whole process; the captured text is re-emitted + at DEBUG level for troubleshooting. + """ + from gradio_client import Client + + logger.info("Connecting to Qwen API at %s...", url) + old_stdout = sys.stdout + captured = io.StringIO() + sys.stdout = captured + try: + try: + client = Client(url, httpx_kwargs={"timeout": config.API_TIMEOUT}) + except TypeError: + # Older gradio_client versions don't support httpx_kwargs. + client = Client(url) + if clone: + self.clone_client = client + self.clone_api_info = self._load_api_info(client) + else: + self.client = client + self.api_info = self._load_api_info(client) + finally: + sys.stdout = old_stdout + usage_info = captured.getvalue().strip() + if usage_info: + logger.debug("Gradio client output for %s:\n%s", url, usage_info) + logger.info("Connected to Qwen API") + + @staticmethod + def _load_api_info(client) -> Dict[str, Any]: + """Load available API metadata from the Gradio app.""" + try: + return client.view_api(return_format="dict") + except Exception as exc: + logger.warning("Unable to read API metadata: %s", exc) + return {} + + def _resolve_api_name(self, *candidates: str, api_info: Optional[Dict[str, Any]] = None) -> str: + """Return the first available api_name from candidate list.""" + info = api_info if api_info is not None else self.api_info + named_endpoints = info.get("named_endpoints", {}) + for candidate in candidates: + if candidate in named_endpoints: + return candidate + return candidates[0] + + def _endpoint_accepts_param(self, api_name: str, param_name: str, + api_info: Optional[Dict[str, Any]] = None) -> bool: + """Check whether endpoint input schema includes the given parameter.""" + info = api_info if api_info is not None else self.api_info + endpoint = info.get("named_endpoints", {}).get(api_name, {}) + parameters = endpoint.get("parameters", []) + return any(parameter.get("parameter_name") == param_name for parameter in parameters) + + # ------------------------------------------------------------------ + # Reference audio transcription (voice clone) + # ------------------------------------------------------------------ + + def transcribe_audio(self, audio_path: str) -> Optional[str]: + """Transcribe reference audio locally using an optional Whisper backend.""" + from .transcribe import transcribe_reference_audio + return transcribe_reference_audio(audio_path) + + # ------------------------------------------------------------------ + # Chunk generation + # ------------------------------------------------------------------ + + def generate_chunk(self, text: str, chunk_num: int) -> Optional[str]: + """Generate one audio chunk; returns its path in the chunks folder. + + The text is split into sub-requests of at most + ``config.CHUNK_SIZE`` words each (the book-level chunker + normally guarantees this already; the split is defense in depth + against pathological input such as a punctuation-free run of + text), and the audio files returned for the sub-requests are + concatenated into one chunk file. + """ + try: + sub_texts = split_into_chunks(text, max_words=config.CHUNK_SIZE) + if not sub_texts: + raise RuntimeError("No text to synthesize") + + output_path: Optional[Path] = None + with tempfile.TemporaryDirectory(prefix="tts_parts_") as parts_dir, \ + self._chunk_heartbeat(chunk_num): + part_paths = [ + self._generate_sub_request(sub_text, parts_dir, sub_num, + len(sub_texts), chunk_num) + for sub_num, sub_text in enumerate(sub_texts, 1) + ] + if len(part_paths) == 1: + suffix = part_paths[0].suffix or ".wav" + output_path = self._chunk_path(chunk_num, suffix) + shutil.copy2(part_paths[0], output_path) + else: + output_path = self._chunk_path(chunk_num, ".wav") + concat_audio_files(part_paths, output_path) + + logger.debug("Chunk %d generated successfully (%d sub-request(s))", + chunk_num, len(sub_texts)) + return str(output_path) + + except ConversionCancelled: + raise + except Exception as exc: + logger.error("Qwen chunk processing failed for chunk %d: %s", chunk_num, exc) + return None + + def _generate_sub_request(self, text: str, parts_dir: str, sub_num: int, + sub_total: int, chunk_num: int) -> Path: + """Run one API generation for ``text``; returns the downloaded audio.""" + if sub_total > 1: + logger.info("Chunk %d: oversized input split into %d requests " + "(sub-request %d/%d)", chunk_num, sub_total, sub_num, sub_total) + if self.voice_mode == VOICE_MODE_CUSTOM: + result = self._generate_custom_voice(text) + elif self.voice_mode == VOICE_MODE_CLONE: + result = self._generate_voice_clone(text) + else: + raise ValueError(f"Unknown voice mode: {self.voice_mode}") + + if not isinstance(result, (tuple, list)) or not result: + raise RuntimeError("Qwen API returned an invalid result") + + audio_path = result[0] # First element is the audio file path + if not isinstance(audio_path, (str, Path)) or not audio_path: + raise RuntimeError("Qwen API did not return an audio file path") + + source = Path(audio_path) + if not source.exists(): + raise RuntimeError(f"Generated audio file not found: {audio_path}") + + destination = Path(parts_dir) / f"part_{sub_num:02d}{source.suffix or '.wav'}" + shutil.copy2(source, destination) + + return destination + + # ------------------------------------------------------------------ + # API payloads + # ------------------------------------------------------------------ + + def _generate_custom_voice(self, text: str) -> Tuple: + """Generate audio using CustomVoice mode.""" + custom_api = self._resolve_api_name("/run_instruct", "/run_custom_voice", "/generate_custom_voice") + if custom_api == "/run_instruct": + payload = dict( + text=text, + lang_disp=self.language, + spk_disp=speaker_display_name(), + instruct=config.INSTRUCT, + ) + else: + payload = dict( + text=text, + language=self.language, + speaker=config.SPEAKER, + instruct=config.INSTRUCT, + ) + if self._endpoint_accepts_param(custom_api, "model_id_cv"): + payload["model_id_cv"] = CUSTOM_VOICE_MODEL_ID + elif self._endpoint_accepts_param(custom_api, "model_size"): + payload["model_size"] = MODEL_SIZE + + if self._endpoint_accepts_param(custom_api, "seed"): + payload["seed"] = self._seed + + return self.client.predict(**payload, api_name=custom_api) + + def _ref_audio_payload(self) -> Dict[str, Any]: + """Gradio file payload for the reference audio (built once, reused).""" + if self._ref_audio_filedata is None: + from gradio_client import handle_file + self._ref_audio_filedata = handle_file(self.voice_clone_ref_audio) + return self._ref_audio_filedata + + def _generate_voice_clone(self, text: str) -> Tuple: + """Generate audio using Voice Clone mode.""" + if not Path(self.voice_clone_ref_audio).exists(): + raise FileNotFoundError(f"Reference audio not found: {self.voice_clone_ref_audio}") + + if self.clone_client is None: + raise RuntimeError("Voice Clone client is not initialized. Is the Base-model demo running?") + + clone_api = self._resolve_api_name("/run_voice_clone", "/generate_voice_clone", + api_info=self.clone_api_info) + use_xvector = config.XVECTOR_ONLY or not self.voice_clone_ref_text + + if clone_api == "/run_voice_clone": + payload = dict( + ref_aud=self._ref_audio_payload(), + ref_txt=self.voice_clone_ref_text, + use_xvec=use_xvector, + text=text, + lang_disp=self.language, + ) + else: + payload = dict( + ref_audio=self._ref_audio_payload(), + ref_text=self.voice_clone_ref_text, + target_text=text, + language=self.language, + use_xvector_only=use_xvector, + ) + optional_params = { + "model_size": MODEL_SIZE, + "seed": self._seed, + } + for name, value in optional_params.items(): + if self._endpoint_accepts_param(clone_api, name, api_info=self.clone_api_info): + payload[name] = value + + return self.clone_client.predict(**payload, api_name=clone_api) diff --git a/app/converter/clients/speakers.py b/app/converter/clients/speakers.py new file mode 100644 index 0000000..eecd52a --- /dev/null +++ b/app/converter/clients/speakers.py @@ -0,0 +1,57 @@ +"""Qwen3-TTS built-in speaker names and their wire (display) forms.""" + +from typing import Optional + +from .. import config + +# 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", + "serena": "Serena", + "vivian": "Vivian", + "uncle_fu": "Uncle Fu", + "aiden": "Aiden", + "ono_anna": "Ono Anna", + "sohee": "Sohee", + "eric": "Eric", + "dylan": "Dylan", +} + + +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 --voice / + Speaker-picker value into what audiocpp_server expects in the request's + voice field. + """ + return SPEAKER_DISPLAY_NAMES.get((name or "").lower(), name) + + +def is_builtin_speaker(name: Optional[str]) -> bool: + """True when NAME is one of the Qwen3-TTS CustomVoice built-in speakers. + + Matches case-insensitively across the canonical ("Uncle_Fu"), display + ("Uncle Fu") and shorthand ("uncle_fu") forms, so the --voice flag and + the Convert form's Speaker picker resolve to the same set. + """ + if not name: + return False + norm = name.lower().replace("_", " ").replace("-", " ") + return any(norm == speaker.lower().replace("_", " ") + for speaker in QWEN3_TTS_SPEAKERS) + + +def speaker_display_name() -> str: + """Return the display name for the configured custom speaker.""" + return speaker_display_name_for(config.SPEAKER) diff --git a/app/converter/clients/transcribe.py b/app/converter/clients/transcribe.py new file mode 100644 index 0000000..d2db9f1 --- /dev/null +++ b/app/converter/clients/transcribe.py @@ -0,0 +1,54 @@ +"""Optional local Whisper transcription of reference audio.""" + +import logging +from typing import Optional + +logger = logging.getLogger(__name__) + + +def transcribe_reference_audio(audio_path: str, model_name: str = "base") -> Optional[str]: + """Transcribe reference audio locally using an optional Whisper backend. + + The current qwen-tts demo does not expose a transcription endpoint, so + transcription is done client-side when a Whisper package is available. + Returns None if no backend is installed. + """ + for backend in ("faster_whisper", "whisper"): + try: + if backend == "faster_whisper": + from faster_whisper import WhisperModel + model = WhisperModel(model_name, device="cpu", compute_type="int8") + segments, _ = model.transcribe(audio_path) + text = " ".join(seg.text.strip() for seg in segments).strip() + else: + import whisper + model = whisper.load_model(model_name) + result = model.transcribe(audio_path) + text = (result.get("text") or "").strip() + if text: + logger.info("Transcription complete via %s: %s", backend, text) + return text + except ImportError: + continue + except Exception as exc: + logger.warning("%s transcription failed: %s", backend, exc) + logger.warning("No Whisper backend available; transcription skipped.") + return None + + +def whisper_backend_available() -> Optional[str]: + """Return the name of an importable Whisper backend, or None. + + Checks faster_whisper first (preferred), then the openai-whisper + package, without importing the heavy model code: a bare import probe + is enough to tell whether the package is installed in the current + environment. Used by the make_audiocpp_server_json tool to warn when + neither is present (e.g. the wrong conda environment is active). + """ + for backend in ("faster_whisper", "whisper"): + try: + __import__(backend) + except ImportError: + continue + return backend + return None diff --git a/app/converter/converter.py b/app/converter/converter.py index 1abc85c..48d4987 100644 --- a/app/converter/converter.py +++ b/app/converter/converter.py @@ -15,7 +15,7 @@ from typing import Callable, Dict, List, Optional, Tuple from . import audio, chunking, config, cover, extractors from .audio import TrackMeta -from .tts import ( +from .clients import ( BACKENDS, BACKEND_AUDIOCPP, BACKEND_FASTER, @@ -228,7 +228,8 @@ class AudiobookConverter: if backend == BACKEND_FASTER: # The faster backend always voice-clones using a reference voice # configured on the server, so no local reference audio is needed. - self.tts = FasterTTSClient(voice=voice, api_url=api_url, + self.tts = FasterTTSClient(chunks_dir=CHUNKS_FOLDER, + voice=voice, api_url=api_url, quiet=quiet) elif backend == BACKEND_AUDIOCPP: # --voice picks the voice: a built-in speaker name on the @@ -237,13 +238,15 @@ class AudiobookConverter: # 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, + self.tts = AudioCppTTSClient(chunks_dir=CHUNKS_FOLDER, + voice=voice, language=self.language, model_id=model_id, instructions=instructions, request_options=self.request_options, api_url=api_url, quiet=quiet) else: self.tts = QwenTTSClient( + chunks_dir=CHUNKS_FOLDER, voice_mode=voice_mode, voice_clone_ref_audio=voice_clone_ref_audio, voice_clone_ref_text=voice_clone_ref_text, @@ -398,7 +401,7 @@ class AudiobookConverter: self.current_outputs = [] # Start from a clean scratch folder so a previous crash can never # affect this run - audio.cleanup_chunks() + audio.cleanup_chunks(CHUNKS_FOLDER) logger.info("Extracting text...") book = extractors.extract_book(file_path) @@ -476,7 +479,7 @@ class AudiobookConverter: return False finally: # Always cleanup, even on failure or interrupt - audio.cleanup_chunks() + audio.cleanup_chunks(CHUNKS_FOLDER) def _convert_m4b_with_chapters(self, sections, stem: str, start_time: float, meta: Optional[TrackMeta] = None, @@ -519,7 +522,9 @@ class AudiobookConverter: return False output_path = AUDIOBOOKS_FOLDER / f"{stem}.{self.output_format}" - if not audio.combine_chapters_to_m4b(chapter_files, titles, output_path, speed=self.speed, + if not audio.combine_chapters_to_m4b(chapter_files, titles, output_path, + chunks_dir=CHUNKS_FOLDER, + speed=self.speed, meta=meta, cover=cover): return False duration = time.time() - start_time @@ -653,7 +658,9 @@ class AudiobookConverter: successful_chunks, total_chunks) return False - success = audio.combine_chunks(total_chunks, output_path, chunk_results=results, + success = audio.combine_chunks(total_chunks, output_path, + chunk_results=results, + chunks_dir=CHUNKS_FOLDER, speed=speed, output_format=output_format, intermediate=chapter is not None, meta=meta, cover=cover) |