fix: silently failing whisper transcription

2 files changed, 42 insertions, 14 deletions

diff --git a/app/backends/audiocpp/voices.py b/app/backends/audiocpp/voices.py
index b26a0be..d19cb18 100644
--- a/app/backends/audiocpp/voices.py
+++ b/app/backends/audiocpp/voices.py
@@ -2,19 +2,22 @@
 
 import argparse
 from pathlib import Path
-from typing import Dict, List, Optional, Tuple
+from typing import Dict, Optional, Tuple
 
 from backends.common import (PROMPT_TEXT_FILENAME, find_wav_files,
                              read_prompt_text)
-from converter.clients import (transcribe_reference_audio,
-                               whisper_backend_available)
+from converter.clients import (transcribe_reference_audio_detailed,
+                               whisper_backend_problem)
 
 def transcribe_wav_dir(wav_files: list, whisper_model: str,
                        cancel=None) -> Dict[str, str]:
-    """Transcribe each wav file and return a mapping of stem -> transcript.
+    """Transcribe each wav file and return a stem -> transcript mapping.
 
     CANCEL (a ``threading.Event``) is checked between files so the in-TUI
-    task view can stop a long transcription early.
+    task view can stop a long transcription early. A wav that yields no
+    transcript keeps its '' entry, and the per-file warning says WHY
+    (backend missing, import broken, transcribe error, or no speech)
+    instead of a bare 'no transcript' line.
     """
     transcripts: Dict[str, str] = {}
     for wav_file in wav_files:
@@ -23,13 +26,15 @@ def transcribe_wav_dir(wav_files: list, whisper_model: str,
             break
         name = wav_file.stem
         print(f"[INFO] Transcribing {wav_file.name} (voice '{name}')...")
-        text = transcribe_reference_audio(str(wav_file), model_name=whisper_model)
+        text, reason = transcribe_reference_audio_detailed(
+            str(wav_file), model_name=whisper_model)
         if text:
             print(f"[OK] {name}: {text}")
         else:
-            print(f"[WARNING] No transcript for '{name}'; cloning works best "
-                  "with an accurate transcript — consider editing prompt_text "
-                  "by hand before starting the server")
+            print(f"[WARNING] No transcript for '{name}': {reason}")
+            print("[WARNING] Cloning works best with an accurate transcript; "
+                  "consider editing prompt_text by hand before starting "
+                  "the server")
         transcripts[name] = text or ""
     return transcripts
 
@@ -79,11 +84,10 @@ def _transcribe(args: argparse.Namespace, plan: Optional[dict],
               "already transcribed, nothing new to transcribe")
         return existing, False
 
-    if whisper_backend_available() is None:
-        print("[WARNING] Neither faster_whisper nor whisper was found, so "
-              "reference .wav files cannot be transcribed automatically and "
-              "every transcript will be empty.")
-        print("          Install whisper (or faster_whisper) in your "
+    problem = whisper_backend_problem()
+    if problem is not None:
+        print(f"[WARNING] No usable Whisper backend: {problem}")
+        print("          Install faster-whisper (or openai-whisper) in your "
               "audiobook environment to transcribe automatically; otherwise "
               "transcripts must be added by hand (see the warning at the end).")
 
@@ -95,6 +99,17 @@ def _transcribe(args: argparse.Namespace, plan: Optional[dict],
     else:
         transcripts = transcribe_wav_dir(wav_files, args.whisper_model,
                                          cancel=cancel)
+
+    # A failed transcription must never wipe a good transcript: keep the
+    # existing text for any voice whose new transcription came back empty.
+    # A blank prompt_text entry makes the server reject every clone request
+    # for that voice, so overwriting known-good text with '' is always a
+    # regression, whatever went wrong with the re-transcription.
+    for name, text in transcripts.items():
+        if not text.strip() and existing.get(name, "").strip():
+            transcripts[name] = existing[name]
+            print(f"[WARNING] Kept the existing transcript for '{name}' "
+                  "(the new transcription came back empty)")
     return transcripts, True
 
 
diff --git a/app/backends/audiocpp/wizard.py b/app/backends/audiocpp/wizard.py
index 91a8522..46b6638 100644
--- a/app/backends/audiocpp/wizard.py
+++ b/app/backends/audiocpp/wizard.py
@@ -600,6 +600,19 @@ def _execute_lanes(settings: dict,
             transcripts, write_prompt = {}, False
         state["transcripts"] = transcripts
         state["write_prompt"] = write_prompt
+        # A blank transcript means the affected clone voices cannot work:
+        # transcript-conditioned families (Qwen3-TTS Base ICL) reject every
+        # request without it. Report the step as failed instead of a silent
+        # [OK]; the setup continues (warn-and-continue) and still writes
+        # server.json and prompt_text.
+        blank = sorted(name for name, text in transcripts.items()
+                       if not text.strip())
+        if blank:
+            print("[ERROR] No transcript for: " + ", ".join(blank))
+            print("[ERROR] Those clone voices will NOT work until prompt_text "
+                  "carries an accurate transcript for each one; see the "
+                  "summary below for how to fix them by hand.")
+            return 1
         return 0
 
     def write(emit, cancel):