Verified empirically (gitea #7): the neosun wrapper's \`instruct\`
field routes through CosyVoice's inference_instruct2 which expects
a Chinese template. English directives produce Chinese-phonemed
speech of the directive itself — the model is stuck in Chinese
context regardless of the directive language.
XML inline tags (<angry>, <sad>, etc.) bypass that path entirely
and work cleanly for English emotion control.
README updated:
- Added Gotcha section entry explaining the instruct trap with
the specific failure mode and the byte-identical-response
evidence
- Smoke-test recipes (steps 4 + 5) reworked to use <angry>...</angry>
in the input instead of an English instruct value
- "Emotion / style control" table now marks instruct as Chinese-
only with ⚠️ and XML tags as ✅
The original smoke-test curl used voice="default" which doesn't
exist — the neosun wrapper ships zero preset voices, and the
built-in SFT speakers (中文男/女 etc.) are not surfaced. Calling
/v1/audio/speech with any unregistered voice returns a 400 whose
JSON body curl happily writes into the .wav (124-byte phantom).
Replaced the smoke test with the full clone → synthesize flow and
added a gotchas section covering:
- No default voice; /v1/voices/create is mandatory
- Reference audio ≤30s (frontend asserts; longer clips 500 at
synthesis time, not at upload)
- Providing an explicit transcript beats the auto-ASR fallback
- voice_id is the handle, not name
- Both cosyvoice-v3 and cosyvoice-v2 ship in the image
Also documented streaming: available via /api/tts with stream=true
(~150ms TTFB), NOT on /v1/audio/speech. Clarified field-name
differences between the OpenAI-compat and native endpoints in a
table. No WebSocket / SSE in this wrapper despite upstream support.
Two new speech stacks on irv-ml1, both on the /worktank/<stack>/
pattern, no tnet (irv-ml1 is local-endpoints-only for now).
parakeet — ASR via Shadowfita/parakeet-tdt-0.6b-v2-fastapi:
- docker buildx git context pinned to SHA 31c5652; no source
vendored. Rebuild on SHA bump.
- GPU-capable FastAPI + Silero VAD + WS streaming.
- API: POST /transcribe, WS /ws/transcribe, GET /healthz. Not the
literal OpenAI `/v1/audio/transcriptions` path — note in README.
- HF cache at /worktank/parakeet/models/ (excluded from restic).
- Build ~158s first time; steady-state start ~40s.
cosyvoice — TTS via neosun/cosyvoice:v1.3.2 shipping
Fun-CosyVoice3-0.5B-2512 (CosyVoice 3, chosen over v2 for the
expanded 5,000-hour instruction-following data covering emotions,
speed, tones, dialects, accents, role-playing; ~150ms streaming
TTFB matches v2). API: /v1/audio/speech (OpenAI drop-in),
/v1/voices/create (cloning), /health.
- Host port 8190 (container 8188; host 8188 already taken by comfyui).
- /worktank/cosyvoice/{voices,input,output}/; voices include in
restic (precious — reproducing a clone needs the original ref
audio), input+output excluded (scratch).
- Model weights (~2-3 GB) live inside image layer; re-download on
tag bump, persist across `compose up -d`.
Both healthy on first deploy.
vh