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.

playbooks/deploy-omnivoice.yaml

@@ -0,0 +1,110 @@
+# Deploy OmniVoice (https://github.com/k2-fsa/OmniVoice) to irv-ml1, GPU 0
+# (RTX 3090). Apache-2.0 zero-shot multilingual voice-cloning TTS, served
+# via upstream's own Gradio demo (no FastAPI wrapper).
+#
+# Builds the image locally from stacks/omnivoice/Dockerfile (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 run is slow: ~5-10 min docker build + a one-time HF weight pre-warm
+# (k2-fsa/OmniVoice) on first container start (entrypoint.sh). The wait loop
+# below allows up to ~20 min for build-then-up + pre-warm.
+#
+# Usage:
+#   scripts/elway irv-ml1 --playbook playbooks/deploy-omnivoice.yaml
+#
+# Idempotent — every step is creates-/when-gated; rerun is safe.
+
+vars:
+  compose_dir: /opt/docker/compose/omnivoice
+  cache_dir: /worktank/omnivoice/hf_cache
+  voices_dir: /worktank/omnivoice/voices
+  host_port: "8199"
+
+steps:
+  # ── host-side dirs ──────────────────────────────────────────────────
+
+  - name: Ensure /worktank/omnivoice root exists (one-time, sudo)
+    shell: mkdir -p /worktank/omnivoice
+    sudo: true
+    creates: /worktank/omnivoice
+
+  - name: Chown /worktank/omnivoice to lkraven
+    shell: chown lkraven:lkraven /worktank/omnivoice
+    sudo: true
+    when: '[ "$(stat -c %U /worktank/omnivoice)" != lkraven ]'
+
+  - name: Ensure cache dir exists
+    shell: mkdir -p {{ cache_dir }}
+    creates: "{{ cache_dir }}"
+
+  - name: Ensure voices dir exists
+    shell: mkdir -p {{ voices_dir }}
+    creates: "{{ voices_dir }}"
+
+  - name: Ensure compose dir exists
+    shell: mkdir -p {{ compose_dir }}
+    creates: "{{ compose_dir }}"
+
+  # ── deploy build context (compose, dockerfile, entrypoint, env) ───────
+
+  - name: Upload compose.yaml
+    upload:
+      src: stacks/omnivoice/compose.yaml
+      dest: "{{ compose_dir }}/compose.yaml"
+      mode: "0644"
+
+  - name: Upload Dockerfile
+    upload:
+      src: stacks/omnivoice/Dockerfile
+      dest: "{{ compose_dir }}/Dockerfile"
+      mode: "0644"
+
+  - name: Upload entrypoint.sh
+    upload:
+      src: stacks/omnivoice/entrypoint.sh
+      dest: "{{ compose_dir }}/entrypoint.sh"
+      mode: "0755"
+
+  - name: Seed .env from template (only if absent)
+    upload:
+      src: stacks/omnivoice/.env.example
+      dest: "{{ compose_dir }}/.env"
+      mode: "0644"
+    when: "[ ! -f {{ compose_dir }}/.env ]"
+
+  # ── build + bring up ────────────────────────────────────────────────
+
+  - name: docker compose build (~5-10 min first time; cached after)
+    shell: |
+      set -o pipefail
+      cd {{ compose_dir }} && docker compose build --progress=plain 2>&1 \
+        | grep -vE '^#[0-9]+ [0-9.]+ (Downloading|Collecting|Requirement|Using cached|Installing collected|Successfully (installed|built)|━|Resolved|Prepared|Built)'
+
+  - name: docker compose up -d
+    shell: cd {{ compose_dir }} && docker compose up -d
+
+  - name: Wait for the Gradio UI to respond (allow ~20 min for weight pre-warm)
+    shell: |
+      for i in $(seq 1 240); do
+        curl -sf -o /dev/null --max-time 3 http://localhost:{{ host_port }}/ && exit 0
+        sleep 5
+      done
+      exit 1
+    changed_when: "false"
+
+verify:
+  - name: Gradio UI returns 200
+    shell: curl -sf -o /dev/null http://localhost:{{ host_port }}/
+    changed_when: "false"
+
+  - name: Container is running
+    shell: docker inspect omnivoice --format '{{.State.Status}}' | grep -q running
+    changed_when: "false"
+
+  - name: Container is healthy (or still starting weights)
+    shell: |
+      s=$(docker inspect omnivoice --format '{{.State.Health.Status}}' 2>/dev/null)
+      echo "health: $s"; [ "$s" = healthy ] || [ "$s" = starting ]
+    changed_when: "false"

stacks/omnivoice/.env.example

@@ -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

stacks/omnivoice/Dockerfile

@@ -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"]

stacks/omnivoice/README.md

@@ -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).

stacks/omnivoice/compose.yaml

@@ -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}

stacks/omnivoice/entrypoint.sh

@@ -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 "$@"