reference_id=<name> resolves against the DIRECTORY references/<name>/
(audio + same-basename .lab), not a flat references/<name>.wav. Voices
were staged flat with the per-name dirs left empty, so every
reference_id resolved to nothing and Fish fell back to its default
speaker — every dropdown voice produced byte-identical audio (proven:
Abigail == Imogen == no-ref, same text+seed). This was the real "no
accent" root cause, independent of the asset-engine "undefined" select
bug.
Server fix (applied to irv-ml1): populated references/<name>/<name>.wav
+ <name>.lab for all 32 voices; re-test confirms Imogen/Eleanor/
Beatrice/Abigail/no-ref now all distinct.
Durable hardening + record correction:
- playbook: normalize-layout step (flat <name>.wav -> nested dir, cp -u
idempotent, when-gated on count mismatch) + an A/B verify gate that
hard-fails the deploy if two reference_ids yield identical output.
- services.yaml: correct the reference_id resolution doc (dir + .lab,
not flat wav).
- README + persistent-memory: correct the "reference_id-by-name is THE
working path, verified" claim — it was a no-op until this fix; the
prior ECAPA 0.79 result came through the inline base64 path.
After getting fish-s2 finally healthy on attempt #5, the playbook's
verify still failed because /v1/audio/voices doesn't exist. Discovery:
the Fish wrapper has a custom API surface, not OpenAI-compatible.
Real endpoints:
POST /v1/tts — synthesis (text body, optional `references`
field for voice cloning, returns audio/wav)
GET /v1/health — liveness (used by Docker healthcheck)
GET /heartbeat — alternate liveness signal
GET / — Swagger Editor UI for the OpenAPI spec
No /v1/audio/speech, /v1/audio/voices, /v1/models — those return 404.
Updated:
* Playbook verify — replaced the JSON-shape /v1/audio/voices check
with a POST /v1/tts smoke that asserts a real RIFF WAV comes back.
* README API section — replaced the OpenAI-compat examples with
Fish's actual {"text":"...","references":[...]} body shape.
* README disk footprint — corrected ~9 GB → ~11 GB (codec.pth was
larger than I estimated; 1.9 GB + 9 GB safetensors).
* README Lessons learned section — recorded the 5-iteration deploy
story so the next time we touch a Fish-style upstream we don't
re-walk the dockerfile / target / pre-pull / API-shape traps.
Third deploy attempt got past the build but crashlooped at container
start: Fish's start_server.sh validates checkpoints/s2-pro/ exists
and exits cleanly (rc=0) if missing — no auto-download, no helpful
message. /worktank/fish-s2/checkpoints/ was empty, so the container
exited every ~52s under restart policy.
Added an idempotent pre-pull step using the same one-shot
python:3.12-slim + huggingface_hub.snapshot_download + hf_transfer
pattern we used for the Qwen 3.6 GGUFs earlier today. Pulls the 9
relevant files (~11 GB total: codec.pth + 2 safetensors shards +
config + tokenizer/template) directly into the bind-mount at
/worktank/fish-s2/checkpoints/s2-pro/ — gated by `creates:` on
codec.pth so the pre-pull step is a no-op on reruns.
~83 s wall-clock for the 11 GB pull on first deploy.
Adds the three premier 2026 TTS releases we missed during the original
fleet build-out (early April), all licensed for self-host:
* Fish Audio S2-Pro (port 8195, GPU 1 / A6000) — released 2026-03-09.
4B dual-AR (Slow + Fast) trained on 10M+ hours / 80+ languages.
Headline: 15,000+ paralinguistic / emotion tags via natural language
([laugh] [whispers] [super happy] etc.) — a step-function over
Chatterbox Turbo's 9 fixed tags. 91.61% paralinguistic win rate on
EmergentTTS-Eval. ~150 ms streaming TTFB, voice cloning, MIT-style
open. ~17 GB VRAM.
* Voxtral TTS (port 8197, GPU 1 / A6000) — Mistral, released 2026-03-28.
4B open-weight, 70 ms model latency, 9.7× realtime. 68.4% blind A/B
win rate vs ElevenLabs Flash v2.5 in cloning. 8 languages
(EN/FR/DE/ES/IT/PT/NL/HI). Served via vLLM-Omni (Mistral's partner
serving stack) — published Docker image, no local build. ~16 GB VRAM.
CC BY-NC license — personal/research use only; flagged in README.
* Kyutai TTS (port 8198, GPU 0 / 3090) — kyutai/tts-1.6b-en_fr.
Trained on 2.5M hours from the Moshi/Mimi team. Claimed 220 ms in
solo setup, 32 simultaneous streams under 350 ms on L40. Kyutai's
official deploy is Rust + websockets only; using NillPointer's
community OpenAI-compat wrapper to bridge to /v1/audio/speech so
it slots into the same bench harness. ~4-6 GB VRAM.
Each stack: compose.yaml (build context, env, volumes, healthcheck,
homepage label), .env.example (all tunables documented), README.md
(why it exists, headline numbers, API, deploy + hardware notes).
Playbooks at playbooks/deploy-{fish-s2,voxtral,kyutai-tts}.yaml are
idempotent in the same shape as the existing deploy-vibevoice /
deploy-chatterbox playbooks.
Port allocations on irv-ml1 after this lands: 8188 ComfyUI, 8190
CosyVoice, 8191 Qwen3-TTS, 8192 IndexTTS-2, 8193 Kokoro, 8194
VibeVoice, 8195 Fish, 8196 Chatterbox, 8197 Voxtral, 8198 Kyutai,
8765 Parakeet ASR.
vh