feat: improve streaming voice clone service stability

examples/run_voice_clone_service.sh

@@ -0,0 +1,84 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+usage() {
+  cat <<'EOF'
+Usage: run_voice_clone_service.sh [options] [-- extra-args]
+
+Options:
+  -g, --gpu ID           GPU id to use (default: 4)
+  -p, --port PORT        Port to bind (default: 8001)
+  -m, --model MODEL      Model id or path (default: Qwen/Qwen3-TTS-12Hz-1.7B-Base)
+  -s, --voice-store DIR  Voice store directory (default: <repo>/voice_store)
+  -d, --device DEVICE    Device for device_map (default: cuda:0)
+      --ip IP            Bind IP (default: 0.0.0.0)
+  -h, --help             Show this help
+
+Env overrides: GPU_ID, PORT, MODEL, VOICE_STORE, DEVICE, IP, PYTHON_BIN
+EOF
+}
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+PROJECT_ROOT="$(cd "${SCRIPT_DIR}/.." && pwd)"
+
+GPU_ID="${GPU_ID:-4}"
+PORT="${PORT:-8001}"
+MODEL="${MODEL:-Qwen/Qwen3-TTS-12Hz-1.7B-Base}"
+VOICE_STORE="${VOICE_STORE:-${PROJECT_ROOT}/voice_store}"
+DEVICE="${DEVICE:-cuda:0}"
+IP="${IP:-0.0.0.0}"
+PYTHON_BIN="${PYTHON_BIN:-python3}"
+
+EXTRA_ARGS=()
+while [[ $# -gt 0 ]]; do
+  case "$1" in
+    -g|--gpu)
+      GPU_ID="$2"
+      shift 2
+      ;;
+    -p|--port)
+      PORT="$2"
+      shift 2
+      ;;
+    -m|--model)
+      MODEL="$2"
+      shift 2
+      ;;
+    -s|--voice-store)
+      VOICE_STORE="$2"
+      shift 2
+      ;;
+    -d|--device)
+      DEVICE="$2"
+      shift 2
+      ;;
+    --ip)
+      IP="$2"
+      shift 2
+      ;;
+    -h|--help)
+      usage
+      exit 0
+      ;;
+    --)
+      shift
+      EXTRA_ARGS+=("$@")
+      break
+      ;;
+    *)
+      EXTRA_ARGS+=("$1")
+      shift
+      ;;
+  esac
+done
+
+mkdir -p "$VOICE_STORE"
+
+CUDA_VISIBLE_DEVICES="$GPU_ID" "$PYTHON_BIN" \
+  "$SCRIPT_DIR/voice_clone_service.py" \
+  --device "$DEVICE" \
+  --ip "$IP" \
+  --port "$PORT" \
+  --model "$MODEL" \
+  --voice-store "$VOICE_STORE" \
+  "${EXTRA_ARGS[@]}"

examples/run_voice_clone_stream_service.sh

@@ -0,0 +1,84 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+usage() {
+  cat <<'EOF'
+Usage: run_voice_clone_stream_service.sh [options] [-- extra-args]
+
+Options:
+  -g, --gpu ID           GPU id to use (default: 4)
+  -p, --port PORT        Port to bind (default: 8002)
+  -m, --model MODEL      Model id or path (default: Qwen/Qwen3-TTS-12Hz-1.7B-Base)
+  -s, --voice-store DIR  Voice store directory (default: <repo>/voice_store)
+  -d, --device DEVICE    Device for device_map (default: cuda:0)
+      --ip IP            Bind IP (default: 0.0.0.0)
+  -h, --help             Show this help
+
+Env overrides: GPU_ID, PORT, MODEL, VOICE_STORE, DEVICE, IP, PYTHON_BIN
+EOF
+}
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+PROJECT_ROOT="$(cd "${SCRIPT_DIR}/.." && pwd)"
+
+GPU_ID="${GPU_ID:-5}"
+PORT="${PORT:-8002}"
+MODEL="${MODEL:-Qwen/Qwen3-TTS-12Hz-1.7B-Base}"
+VOICE_STORE="${VOICE_STORE:-${PROJECT_ROOT}/voice_store}"
+DEVICE="${DEVICE:-cuda:0}"
+IP="${IP:-0.0.0.0}"
+PYTHON_BIN="${PYTHON_BIN:-python3}"
+
+EXTRA_ARGS=()
+while [[ $# -gt 0 ]]; do
+  case "$1" in
+    -g|--gpu)
+      GPU_ID="$2"
+      shift 2
+      ;;
+    -p|--port)
+      PORT="$2"
+      shift 2
+      ;;
+    -m|--model)
+      MODEL="$2"
+      shift 2
+      ;;
+    -s|--voice-store)
+      VOICE_STORE="$2"
+      shift 2
+      ;;
+    -d|--device)
+      DEVICE="$2"
+      shift 2
+      ;;
+    --ip)
+      IP="$2"
+      shift 2
+      ;;
+    -h|--help)
+      usage
+      exit 0
+      ;;
+    --)
+      shift
+      EXTRA_ARGS+=("$@")
+      break
+      ;;
+    *)
+      EXTRA_ARGS+=("$1")
+      shift
+      ;;
+  esac
+done
+
+mkdir -p "$VOICE_STORE"
+
+CUDA_VISIBLE_DEVICES="$GPU_ID" "$PYTHON_BIN" \
+  "$SCRIPT_DIR/streaming/voice_clone_stream_service.py" \
+  --device "$DEVICE" \
+  --ip "$IP" \
+  --port "$PORT" \
+  --model "$MODEL" \
+  --voice-store "$VOICE_STORE" \
+  "${EXTRA_ARGS[@]}"

examples/streaming/README.md

@@ -0,0 +1,72 @@
+# Voice Clone Streaming Service (SSE)
+
+# Streaming voice cloning service
+
+This service streams voice-clone audio chunks as Server-Sent Events (SSE) or binary PCM.
+
+Files live under `examples/streaming/`.
+
+Constraints
+- Single-sample only (one `text`, one `voice_id` per request).
+- Base 12Hz models only (e.g. `Qwen/Qwen3-TTS-12Hz-1.7B-Base`).
+- SSE payload is base64-encoded `int16` PCM audio.
+
+Start
+```bash
+./Qwen3-TTS/examples/run_voice_clone_stream_service.sh
+```
+
+Test page
+```text
+http://localhost:8002/
+```
+
+Request example
+```bash
+curl -N -X POST \
+  -H "Content-Type: application/json" \
+  -d '{
+    "voice_id": "YOUR_VOICE_ID",
+    "text": "Hello, this is streaming TTS.",
+    "chunk_size": 8,
+    "left_context_size": 25
+  }' \
+  http://localhost:8002/voice/stream
+```
+
+Binary PCM stream
+```text
+POST /voice/stream_pcm
+
+Binary protocol:
+- 12-byte header:
+  - magic: "QTS1" (4 bytes)
+  - sample_rate: uint32 LE
+  - sample_width_bytes: uint16 LE (2 for int16)
+  - channels: uint16 LE (1)
+- Then repeated frames:
+  - frame_len_bytes: uint32 LE
+  - raw PCM int16 bytes
+- End of stream: frame_len_bytes = 0
+```
+
+Model switching
+```bash
+curl -X POST \
+  -H "Content-Type: application/json" \
+  -d '{"model_id":"Qwen/Qwen3-TTS-12Hz-0.6B-Base","safe_mode":true}' \
+  http://localhost:8002/model/load
+```
+
+Available models from the voice store
+```bash
+curl http://localhost:8002/models
+```
+
+Notes
+- Lower `chunk_size` yields faster first audio but more overhead.
+- `left_context_size` improves continuity at chunk boundaries.
+- FlashAttention is enabled by default; use `--no-flash-attn` to disable.
+- TF32 matmul/cudnn is enabled by default on CUDA; use `--no-tf32` to disable.
+- Model switching returns HTTP 409 if a stream is active.
+- Safe mode disables sub-talker sampling for stability.