Two engine-level bugs made custom voices unusable in different ways.
Stale speaker embeddings (voice clone):
A .pt embedding is a cache baked from a specific pairing of reference audio
and reference transcript. Re-recording or re-designing a voice replaces those
sources but leaves the old .pt in place, and generate_voices.py pointed at it
unconditionally — so the server kept cloning from an embedding whose audio
tokens no longer matched the transcript stored alongside them. Confirmed on
73 of 155 voices: generation ignored the requested text entirely, emitting
short unrelated filler ("Thank you.") or fragments of the previous reference
transcript. Voices that had never been re-recorded were unaffected, which is
why this looked specific to designed voices.
Now an embedding older than its reference audio/transcript is treated as
invalid and removed so the server recomputes it. Verified: 73 embeddings
regenerated, and voices that previously returned unrelated text now transcribe
back to exactly the requested input.
Unregistered Voice Design voices:
For the VoiceDesign model a voice's identity IS its instruct prompt, but
designed voices were never written into voicedesign_voices.json, so the server
only knew its 8 bundled presets. resolve_voice() silently substituted the first
one, answering requests for a German male character with 'vd_british_male' —
the source of the apparent gender flips between takes.
- generate_voices.py now mirrors designed voices into the VoiceDesign registry
(it runs in the clone container, which is the one with the voice library
mounted; /config is shared with the VoiceDesign container).
- The VoiceDesign server hot-reloads its registry, matching what the clone
server already did, so voices designed while it is running resolve without
a restart.
- An unknown voice no longer becomes a different one: if the request carries
its own instruct that is used, otherwise it is a 404 rather than a silent
substitution.
- A request instruct is now combined with the voice's registered instruct
instead of replacing it. Previously any line carrying an emotion discarded
the character's identity and re-rolled a voice from a few words of
direction, which made a character drift between lines.
- Per-voice temperature/top_p/top_k are honored. Deliberately no seed:
generate_voice_design() takes no seed parameter, so designed voices cannot
be pinned that way — consistency comes from low temperature/top_p.
Verified end to end: three consecutive takes of the same designed voice now
hold 86-91 Hz median F0 (was flipping register between takes), and emotional
lines stay within 84-86 Hz instead of losing the character entirely.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
POST /voice-seed — writes or clears the seed for a named voice in
voices.json directly from the server, enabling the voice creator GUI
to save the chosen seed with one click.
Also updates find_best_seed.py to support --all-voices, per-voice
subdirectories, and the mixed DE/EN test sentence.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The VoiceClone server was using non_streaming_mode=False, a mode designed
for streaming LLM->TTS pipelines. In that mode only one text token enters
the model's KV cache during prefill; the rest feed via trailing_text_hiddens
at one step per codec frame. For a 54-word paragraph this provides only ~4s
of text guidance for ~18s of speech — 77% generated with no text conditioning.
Without text context the model free-runs and drifts, sometimes changing gender.
Fix: switch to non_streaming_mode=True (already the default for VoiceDesign
and CustomVoice) so the full text is in the prefill throughout generation.
Also lower default temperature 0.9->0.8 and add top_p=0.9 to reduce
accumulated sampling noise over long runs. Temperature, top_k, and top_p
are now configurable per voice in voices.json.
- patches/openai_server.patch: updated for new upstream HEAD; both streaming
(WAV/PCM) and non-streaming (MP3) paths now use non_streaming_mode=True
- config/run_server.py: align warmup call to non_streaming_mode=True
- README.md: bump image tags v4->v5, add changelog section
- Version: v5
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- run_server.py: warm up CUDA graphs at server startup via lifespan
event so the first real request does not pay the 7-8s graph-
compilation penalty; uses modern lifespan API instead of deprecated
on_event
- generate_voices.py: set chunk_size=4 per voice so streaming clients
receive first audio after ~333ms instead of ~1s
- docker-compose.yml: add --max-seq-len 2048 (halves static KV cache,
reduces VRAM and graph-capture time)
- config/benchmark_api.py: new script to measure TTFA, RTF and speed
against the live API endpoint
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Run Qwen3-TTS with CUDA graph acceleration on NVIDIA DGX Spark
(ARM64 / SM 121 / CUDA 13) as an OpenAI-compatible TTS API server.
- Dockerfile targeting nvidia/cuda:13.0.2-base-ubuntu24.04 with ARM64 cu130 PyTorch wheels
- Patch for max-seq-len support to handle long reference audio without crashes
- OpenWebUI + SillyTavern compatible API endpoints (/v1/models, /v1/audio/voices, /speakers)
- Voice management: auto-generate voices.json from speaker reference audio files
- Auto-transcription helper script for generating reference text from audio
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
