feat(omnivoice): new TTS stack — k2-fsa/OmniVoice on irv-ml1 3090

Zero-shot, massively-multilingual (600+ language) voice-cloning + voice-design
TTS (diffusion-LM, Apache-2.0). No official image, so a thin CUDA container
around the pip package running upstream's own Gradio demo (no FastAPI wrapper).
Pinned to GPU 0 (3090) — the A6000 is ComfyUI-exclusive — port 8199. Built +
verified live on irv-ml1 (Gradio 200, container healthy). Surface is the Gradio
UI + Gradio API, NOT OpenAI-compat /v1/audio/speech (wrap later if asset-engine
should consume it). deploy-omnivoice.yaml builds local + verifies.
This commit is contained in:
vh
2026-06-18 22:25:54 -07:00
parent 715a68bee7
commit 984b72757f
6 changed files with 319 additions and 0 deletions
+20
View File
@@ -0,0 +1,20 @@
# OmniVoice (k2-fsa/OmniVoice) — irv-ml1 stack tunables.
# Copy to .env on the host (/opt/docker/compose/omnivoice/.env). The deploy
# playbook seeds .env from this template on first run only.
# Host port (container always listens on 8001). 8199 is free in the irv-ml1
# audio range (8190-8198 + 8765/919x taken; 8201 reserved for voxtral).
OMNIVOICE_PORT=8199
OMNIVOICE_BIND=0.0.0.0
# GPU: device 0 = RTX 3090 on irv-ml1 (device 1 / A6000 is ComfyUI-exclusive).
# OmniVoice runs in <5 GB; the 3090 had ~18 GB free.
OMNIVOICE_GPU_DEVICES=0
# Image tag + optional upstream pin (empty = latest omnivoice on PyPI).
OMNIVOICE_TAG=latest
OMNIVOICE_VERSION=
# Persistent HF weight cache + reference-voice staging on /worktank.
OMNIVOICE_CACHE_DIR=/worktank/omnivoice/hf_cache
OMNIVOICE_VOICES_DIR=/worktank/omnivoice/voices
+54
View File
@@ -0,0 +1,54 @@
# syntax=docker/dockerfile:1.6
#
# OmniVoice (k2-fsa/OmniVoice) — zero-shot, massively-multilingual (600+
# languages) voice-cloning + voice-design TTS, diffusion-LM architecture,
# Apache-2.0. Upstream ships a pip package + its own Gradio demo
# (`omnivoice-demo`); there's no official image, so we build a thin CUDA
# container around the pip package and run its Gradio server directly.
# Unlike index-tts we DON'T need a FastAPI wrapper — OmniVoice serves itself.
ARG CUDA_BASE=nvidia/cuda:12.8.0-cudnn-runtime-ubuntu22.04
FROM ${CUDA_BASE}
ENV DEBIAN_FRONTEND=noninteractive \
PIP_ROOT_USER_ACTION=ignore \
PYTHONUNBUFFERED=1 \
PATH="/opt/venv/bin:${PATH}"
RUN apt-get update \
&& apt-get install -y --no-install-recommends \
python3.10 python3.10-venv python3-pip \
git ffmpeg libsndfile1 \
ca-certificates wget \
&& rm -rf /var/lib/apt/lists/* \
&& python3.10 -m venv /opt/venv
# CUDA 12.8 torch wheels (per OmniVoice's documented install line).
RUN pip install --no-cache-dir \
torch==2.8.0 torchaudio==2.8.0 \
--index-url https://download.pytorch.org/whl/cu128
# OmniVoice from PyPI (+ huggingface_hub for the entrypoint weight pre-warm).
# Optional reproducible pin via the OMNIVOICE_VERSION build arg (empty=latest).
ARG OMNIVOICE_VERSION=
RUN pip install --no-cache-dir "omnivoice${OMNIVOICE_VERSION:+==${OMNIVOICE_VERSION}}" huggingface_hub
# Fail the build loudly if the console script name isn't what we expect,
# rather than crash-loop at runtime. Logs the actual omni* entrypoints.
RUN echo "omni console scripts:" && (ls /opt/venv/bin | grep -i omni || true) \
&& command -v omnivoice-demo >/dev/null \
|| { echo "ERROR: omnivoice-demo CLI not found after install"; exit 1; }
WORKDIR /app
COPY entrypoint.sh /usr/local/bin/entrypoint.sh
RUN chmod +x /usr/local/bin/entrypoint.sh
EXPOSE 8001
# Gradio serves HTML at / — 200 once the UI is up (weights load lazily on
# first synth; the entrypoint pre-warms them). Generous start period.
HEALTHCHECK --interval=30s --timeout=10s --start-period=900s --retries=3 \
CMD wget -q -O /dev/null http://127.0.0.1:8001/ || exit 1
ENTRYPOINT ["/usr/local/bin/entrypoint.sh"]
CMD ["omnivoice-demo", "--ip", "0.0.0.0", "--port", "8001"]
+58
View File
@@ -0,0 +1,58 @@
# OmniVoice
[k2-fsa/OmniVoice](https://github.com/k2-fsa/OmniVoice) — zero-shot,
massively-multilingual (**600+ languages**) voice-cloning + voice-design
TTS from the Next-gen Kaldi / k2-fsa team. Diffusion-LM architecture,
RTF as low as ~0.025 (≈40× real-time). **Apache-2.0** — commercially clean
(unlike Voxtral's CC BY-NC).
## What it does
| Capability | Notes |
|---|---|
| Zero-shot voice cloning | Clone from a short reference clip |
| Voice **design** | Synthesize a voice from attributes (gender, age, pitch, accent, whisper, …) — no reference needed |
| 600+ languages | Broadest coverage of any zero-shot TTS |
| Fine control | Non-verbal symbols + pronunciation correction |
## How it's served
Upstream ships **its own Gradio demo** (`omnivoice-demo`), so this stack
just runs that — no custom wrapper. That means the surface is the **Gradio
UI + Gradio API**, *not* an OpenAI-compatible `/v1/audio/speech` endpoint.
- UI: `http://10.100.79.3:8199/`
- Programmatic: the Gradio API under `/gradio_api/` (or `/config` to
introspect). If you later want OpenAI-compat for asset-engine, add a thin
FastAPI wrapper like [`stacks/index-tts/app.py`](../index-tts/app.py).
## Placement
- **irv-ml1, GPU 0 (RTX 3090)** — pinned via `OMNIVOICE_GPU_DEVICES=0`.
The A6000 (device 1) is ComfyUI-exclusive after the 2026-06-18 VRAM
consolidation. OmniVoice fits in <5 GB; the 3090 had ~18 GB free.
- Port **8199** (8001 inside the container).
## Deploy
```bash
scripts/elway irv-ml1 --playbook playbooks/deploy-omnivoice.yaml
```
Builds the image locally (CUDA 12.8 + torch 2.8.0 + `omnivoice` from PyPI),
stages the build context under `/opt/docker/compose/omnivoice/`, brings it
up, and waits for the Gradio UI on `:8199`. First boot is slow: ~5-10 min
docker build + a one-time HF weight pre-warm (`k2-fsa/OmniVoice`, entrypoint
pre-download into `${OMNIVOICE_CACHE_DIR}`).
## Tunables
All in `.env` (see `.env.example`): `OMNIVOICE_PORT`, `OMNIVOICE_GPU_DEVICES`,
`OMNIVOICE_TAG`, `OMNIVOICE_VERSION` (optional PyPI pin),
`OMNIVOICE_CACHE_DIR`, `OMNIVOICE_VOICES_DIR`. Drop reference WAV/FLAC into
`/worktank/omnivoice/voices/` to stage cloning sources.
## Footprint
- **Disk**: HF weight cache under `/worktank/omnivoice/hf_cache`.
- **VRAM**: <5 GB (docs cite 4 GB+ GPUs).
+48
View File
@@ -0,0 +1,48 @@
# OmniVoice (k2-fsa/OmniVoice) — zero-shot, massively-multilingual (600+
# language) voice-cloning + voice-design TTS, diffusion-LM, Apache-2.0.
# Served via upstream's own Gradio demo. NOTE: this exposes the Gradio UI
# + Gradio API, NOT an OpenAI-compatible /v1/audio/speech endpoint — wrap
# it later (à la stacks/index-tts/app.py) if asset-engine integration is
# wanted. For now it's a "stand it up and try it" UI.
#
# Build: local image from the Dockerfile in this dir. Weights download
# from HF (k2-fsa/OmniVoice) on first boot into ${OMNIVOICE_CACHE_DIR}.
#
# Pinned to the 3090 (device 0) on irv-ml1 — the A6000 (device 1) is
# ComfyUI-exclusive after the 2026-06-18 VRAM consolidation. OmniVoice
# runs in well under 5 GB; the 3090 had ~18 GB free.
#
# All tunables live in .env — edit that, not this file.
services:
omnivoice:
image: local/omnivoice:${OMNIVOICE_TAG:-latest}
build:
context: .
dockerfile: Dockerfile
args:
OMNIVOICE_VERSION: ${OMNIVOICE_VERSION:-}
container_name: omnivoice
restart: unless-stopped
runtime: nvidia
ports:
- "${OMNIVOICE_BIND:-0.0.0.0}:${OMNIVOICE_PORT}:8001"
environment:
- NVIDIA_VISIBLE_DEVICES=${OMNIVOICE_GPU_DEVICES:-0}
- HF_HOME=/app/hf_cache
volumes:
- ${OMNIVOICE_CACHE_DIR}:/app/hf_cache
- ${OMNIVOICE_VOICES_DIR}:/app/voices
healthcheck:
test: ["CMD-SHELL", "wget -q -O /dev/null http://localhost:8001/ || exit 1"]
interval: 30s
timeout: 10s
retries: 3
# First boot: weight pre-warm download (entrypoint) + CUDA warmup.
start_period: 900s
labels:
- homepage.group=AI Systems
- homepage.name=OmniVoice
- homepage.icon=mdi-account-voice
- homepage.description=Zero-shot multilingual voice-cloning TTS (irv-ml1, 3090)
- homepage.href=http://10.100.79.3:${OMNIVOICE_PORT}
+29
View File
@@ -0,0 +1,29 @@
#!/usr/bin/env bash
# entrypoint.sh — pre-warm OmniVoice weights (k2-fsa/OmniVoice) into the
# persistent HF cache on first run, then exec the Gradio demo.
#
# The demo also auto-downloads on first synth, so the pre-warm is
# best-effort (NON-FATAL): it just makes the first generation fast +
# deterministic and lets the healthcheck come up against a ready model.
set -e
: "${HF_HOME:=/app/hf_cache}"
export HF_HOME
mkdir -p "${HF_HOME}"
MARKER="${HF_HOME}/.omnivoice-prewarmed"
if [ ! -f "${MARKER}" ]; then
echo "[omnivoice] pre-warming k2-fsa/OmniVoice weights into ${HF_HOME} (one-time)…"
if python3 - <<'EOF'
from huggingface_hub import snapshot_download
snapshot_download(repo_id="k2-fsa/OmniVoice")
print("[omnivoice] pre-warm complete")
EOF
then
touch "${MARKER}"
else
echo "[omnivoice] pre-warm failed (non-fatal) — demo will download on first synth"
fi
fi
exec "$@"