wip

video/tts/synth.py

@@ -2,7 +2,7 @@
 
 Reads ``output/<storyboard>/narration-script.json`` (emitted by
 ``dist/preflight.js``) and runs ``Qwen3TTSModel.generate_voice_design`` with
-all cue texts as a single batched list — that way every cue shares the same
+all cue texts as a single batched list: that way every cue shares the same
 model state, which keeps prosody and timbre consistent across cues. Per-cue
 WAVs and an index manifest go to ``output/<storyboard>/audio/`` for the
 recording step (which reads measured cue durations) and the mux step (which
@@ -16,7 +16,7 @@ production runs.
 We use the VoiceDesign sibling of CustomVoice because it accepts a free-form
 voice persona (British accent, narrator register, "no laughter") via the
 ``instruct`` parameter. CustomVoice's preset speakers are all American or
-non-English, and its ``instruct`` is documented for emotion only — it
+non-English, and its ``instruct`` is documented for emotion only. It
 ignored accent directives and bled non-speech tokens (laughter, sighs)
 between cues.
 
@@ -43,7 +43,7 @@ from qwen_tts import Qwen3TTSModel
 
 # Two checkpoints: the design model mints the reference clip in the desired
 # persona; the clone model conditions every cue on that reference's x-vector.
-# Neither CustomVoice nor VoiceDesign support generate_voice_clone — only the
+# Neither CustomVoice nor VoiceDesign support generate_voice_clone. Only the
 # Base checkpoint does.
 DEFAULT_DESIGN_MODEL = "Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign"
 DEFAULT_CLONE_MODEL = "Qwen/Qwen3-TTS-12Hz-1.7B-Base"
@@ -69,8 +69,8 @@ def _file_sha256(path: Path) -> str:
     """Content hash of a file, used to pin cached cues to a reference WAV.
 
     The cue cache keys off the *settings* that produced the reference
-    (instruct/seed/…), but a re-mint of VoiceDesign — or render.sh copying a
-    different storyboard's reference into this audio dir — can swap the actual
+    (instruct/seed/…), but a re-mint of VoiceDesign (or render.sh copying a
+    different storyboard's reference into this audio dir) can swap the actual
     reference waveform out from under those settings. Cloning some cues from
     reference A and others from reference B yields two audibly different
     speakers in one video. Hashing the bytes of the reference that was
@@ -328,7 +328,7 @@ def _resolve_reference(
     sf.write(str(ref_wav_path), ref_audio, ref_sr)
     ref_meta_path.write_text(json.dumps(ref_meta, indent=2))
 
-    # Free the design model before loading the clone model — both are 1.7B,
+    # Free the design model before loading the clone model: both are 1.7B,
     # we don't want them resident at the same time.
     del design_model
     if torch.cuda.is_available():
@@ -357,7 +357,7 @@ def main() -> int:
     voice = script.get("voice")
     if not voice:
         print(
-            f"[synth] {script_path} has no `voice` block — re-run preflight.",
+            f"[synth] {script_path} has no `voice` block. Re-run preflight.",
             file=sys.stderr,
         )
         return 1
@@ -386,7 +386,7 @@ def main() -> int:
     #      (or the user supplies one via --reference-audio).
     #   2. Base + generate_voice_clone(x_vector_only_mode=True) conditions
     #      every cue on the reference's speaker embedding.
-    # Without (2), batched generation drifts timbre across cues — a persona
+    # Without (2), batched generation drifts timbre across cues: a persona
     # prompt anchors style but not identity, so each batch item picks its
     # own voice. The reference WAV is cached so subsequent runs only load
     # the clone model (saves ~20s + 3.4 GB of disk download); when it is
@@ -398,13 +398,13 @@ def main() -> int:
     # Pin the cue cache to this exact reference waveform. A re-mint (or a
     # reference copied in from another storyboard by render.sh) changes these
     # bytes; without this gate the cue cache would keep cloned-from-the-old-
-    # reference WAVs alongside freshly regenerated ones — two voices in one
+    # reference WAVs alongside freshly regenerated ones: two voices in one
     # video. See _file_sha256.
     reference_hash = _file_sha256(ref_wav_path)
 
     index_path = audio_dir / "index.json"
 
-    # Skip generation when the existing audio matches the script — same cue
+    # Skip generation when the existing audio matches the script: same cue
     # texts and same gapBeforeMs values in the same order, AND same synth
     # settings (instruct/language/reference text + reference-WAV hash/model/
     # seed/temperature/top_p). Saves ~30s of GPU time when iterating on
@@ -424,7 +424,7 @@ def main() -> int:
         top_p,
     ):
         print(
-            f"[synth] [{args.storyboard}] cached audio matches the current script — skipping generation",
+            f"[synth] [{args.storyboard}] cached audio matches the current script: skipping generation",
             flush=True,
         )
         return 0