feat: add OmniVoice TTS container, upgrade whisper model, tune qwentts tone

2026-09-10 20:43:33 +02:00
@@ -0,0 +1,22 @@
+#!/usr/bin/env bash
+### Build the omnivoice image. USAGE: ./build-container.sh [cpu|rocm] (default: cpu)
+set -euo pipefail
+
+DEVICE="${1:-cpu}"
+
+case "$DEVICE" in
+    cpu|rocm) ;;
+    *)
+        echo "Usage: $0 [cpu|rocm]" >&2
+        exit 1
+        ;;
+esac
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+
+podman build \
+    --build-arg DEVICE="$DEVICE" \
+    -t "localhost/omnivoice:$DEVICE" \
+    -t localhost/omnivoice:latest \
+    -f "$SCRIPT_DIR/omnivoice.Containerfile" \
+    "$SCRIPT_DIR"
@@ -0,0 +1,8 @@
+#!/bin/bash
+# Starts the FastAPI server; model weights download on first request into /app/models
+set -e
+
+echo "=== OmniVoice Server ==="
+echo
+
+exec uvicorn server:app --host 0.0.0.0 --port 8000 --app-dir /app "$@"
@@ -0,0 +1,42 @@
+### OmniVoice Container (https://github.com/k2-fsa/OmniVoice) — build with ./build-container.sh [cpu|rocm]
+ARG DEVICE=cpu
+ARG BASE_IMAGE_CPU=python:3.12-slim
+ARG BASE_IMAGE_ROCM=docker.io/rocm/pytorch-nightly
+
+FROM ${BASE_IMAGE_CPU} AS base-cpu
+FROM ${BASE_IMAGE_ROCM} AS base-rocm
+
+FROM base-${DEVICE}
+ARG DEVICE
+
+USER root
+EXPOSE 8000
+
+# sox/ffmpeg: needed at runtime for reference-audio format handling
+RUN apt-get update \
+    && apt-get install -y --no-install-recommends sox ffmpeg curl \
+    && apt-get autoremove -y \
+    && apt-get clean -y \
+    && rm -rf /var/lib/apt/lists/*
+
+WORKDIR /app
+
+# CPU-only PyTorch (ROCm base image already ships its own build)
+RUN if [ "$DEVICE" = "cpu" ]; then \
+        pip install --no-cache-dir torch torchaudio --index-url https://download.pytorch.org/whl/cpu; \
+    fi
+
+# Install omnivoice and a minimal API server
+RUN pip install --no-cache-dir omnivoice fastapi uvicorn python-multipart soundfile
+
+# Models download at runtime here — mount as a volume to persist them
+RUN mkdir -p /app/models
+ENV HF_HOME=/app/models
+
+# Copy entrypoint / server script
+COPY entrypoint.sh /app/entrypoint.sh
+COPY server.py /app/server.py
+RUN chmod +x /app/entrypoint.sh
+
+ENTRYPOINT ["/app/entrypoint.sh"]
+CMD []
@@ -0,0 +1,13 @@
+omnivoice.example.com {
+	import gateway_error
+
+	request_body {
+		max_size 100MB
+	}
+
+	reverse_proxy omnivoice:8000
+
+	log {
+		output file /var/log/caddy/omnivoice_access.log
+	}
+}
@@ -0,0 +1,36 @@
+[Unit]
+Description=OmniVoice TTS Server
+After=network-online.target
+Wants=network-online.target
+
+[Container]
+Image=localhost/omnivoice:latest
+ContainerName=omnivoice
+
+# CPU-only container
+#Memory=8g
+#CPUs=4
+
+# ROCm
+AddDevice=/dev/kfd
+AddDevice=/dev/dri/renderD128
+PodmanArgs=--group-add=keep-groups --ipc=host --pids-limit=-1 --security-opt label=disable
+SecurityLabelType=container_runtime_t
+
+# HTTP API
+PublishPort=8001:8000
+
+# Persist downloaded model weights across restarts (adjust host path as needed)
+Volume=/srv/containers/omnivoice/models:/app/models:Z
+
+# Optional: override the default checkpoint
+#Environment=OMNIVOICE_MODEL=k2-fsa/OmniVoice
+
+
+[Service]
+Restart=on-failure
+RestartSec=5
+TimeoutStartSec=300
+
+[Install]
+WantedBy=default.target
@@ -0,0 +1,91 @@
+"""Minimal FastAPI server exposing OmniVoice zero-shot voice cloning (/speech/clone only). https://github.com/k2-fsa/OmniVoice"""
+import io
+import os
+import tempfile
+
+import soundfile as sf
+import torch
+from fastapi import FastAPI, File, Form, HTTPException, UploadFile
+from fastapi.responses import StreamingResponse
+from omnivoice import OmniVoice
+
+MODEL_NAME = os.environ.get("OMNIVOICE_MODEL", "k2-fsa/OmniVoice")
+
+_forced_device = os.environ.get("OMNIVOICE_DEVICE", "").strip()
+if _forced_device:
+    DEVICE = _forced_device
+elif torch.cuda.is_available():  # ROCm builds expose the GPU via the CUDA API too
+    DEVICE = "cuda:0"
+elif getattr(torch, "xpu", None) is not None and torch.xpu.is_available():
+    DEVICE = "xpu"
+else:
+    DEVICE = "cpu"
+# float16 is unsupported on CPU and broken on some XPU/oneDNN SDPA builds.
+DTYPE = torch.float16 if DEVICE.startswith("cuda") else torch.float32
+
+# OmniVoice hardcodes fp16 for its internal ASR sub-model on cuda/xpu, which breaks there; keep it on CPU by default.
+ASR_DEVICE = os.environ.get("OMNIVOICE_ASR_DEVICE", "cpu")
+
+app = FastAPI(title="OmniVoice server")
+
+model: OmniVoice | None = None
+
+
+@app.on_event("startup")
+def load_model():
+    global model
+    print(f"Loading OmniVoice model {MODEL_NAME} on {DEVICE}...")
+    model = OmniVoice.from_pretrained(
+        MODEL_NAME,
+        device_map=DEVICE,
+        dtype=DTYPE,
+        asr_device=ASR_DEVICE,
+    )
+    print(f"OmniVoice model loaded. ASR (ref_text auto-transcribe) on {ASR_DEVICE}.")
+
+
+@app.get("/health")
+def health():
+    return {
+        "status": "ok",
+        "model": MODEL_NAME,
+        "device": DEVICE,
+    }
+
+
+@app.post("/speech/clone")
+def speech_clone(
+    text: str = Form(...),
+    language: str = Form(""),
+    ref_text: str = Form(""),
+    instruct: str = Form(""),
+    speed: float = Form(1.0),
+    ref_audio: UploadFile = File(...),
+):
+    """Clone a voice from ref_audio (ref_text auto-transcribed via Whisper if omitted) and synthesize text with it."""
+    if model is None:
+        raise HTTPException(status_code=503, detail="Model not loaded yet")
+
+    ref_bytes = ref_audio.file.read()
+
+    suffix = os.path.splitext(ref_audio.filename or "")[1] or ".wav"  # generate() needs a file path, not bytes
+    with tempfile.NamedTemporaryFile(suffix=suffix, delete=False) as tmp:
+        tmp.write(ref_bytes)
+        tmp_path = tmp.name
+
+    try:
+        audio = model.generate(
+            text=text,
+            language=language or None,
+            ref_audio=tmp_path,
+            ref_text=ref_text or None,
+            instruct=instruct or None,
+            speed=speed,
+        )
+    finally:
+        os.remove(tmp_path)
+
+    buf = io.BytesIO()
+    sf.write(buf, audio[0], model.sampling_rate, format="WAV")
+    buf.seek(0)
+    return StreamingResponse(buf, media_type="audio/wav")
@@ -36,7 +36,7 @@ Volume=/srv/containers/qwen-tts/models:/app/models:Z
 # `instruct` (Qwen3-TTS has no numeric speed/style knob, only natural-language
 # steering). Only applies to /speech (CustomVoice); /speech/clone has no
 # instruct support.
-#Environment=QWEN_TTS_DEFAULT_INSTRUCT=This is a voice message between two close friends on WhatsApp. Speak in a casual, warm, friendly tone, not formal or professional-sounding.
+Environment=QWEN_TTS_DEFAULT_INSTRUCT=This is a voice message between two close friends on WhatsApp. Speak calmly and naturally at a steady, even pace, like ordinary spoken conversation. Keep the tone warm and casual but restrained: avoid dramatic emphasis, exaggerated excitement, or theatrical intonation. Use natural, evenly spaced pauses between sentences, not rushed or drawn out.


 [Service]
@@ -15,11 +15,11 @@ PodmanArgs=--group-add=keep-groups --ipc=host
 SecurityLabelType=container_runtime_t

 # Whisper configuration
-Environment=WHISPER_MODEL_NAME=small
-Environment=WHISPER_MODEL_FILE=ggml-small.bin
+Environment=WHISPER_MODEL_NAME=large-v3-turbo
+Environment=WHISPER_MODEL_FILE=ggml-large-v3-turbo.bin
 Environment=WHISPER_LANGUAGE=auto
 Environment=WHISPER_TRANSLATE=false
-Environment=WHISPER_DTW=large.v3
+#Environment=WHISPER_DTW=large.v3

 Environment=WHISPER_DETECT_LANGUAGE=false
 Environment=HF_HOME=/app/models/.huggingface