Docker for PocketTTS.cpp

.dockerignore

@@ -0,0 +1,6 @@
+# Keep the Docker build context to only the files copied by Dockerfile.
+*
+
+!export_onnx.py
+!CMakeLists.txt
+!pocket_tts.cpp

CMakeLists.txt

@@ -32,7 +32,7 @@ include(FetchContent)
 set(FETCHCONTENT_QUIET OFF)
 
 # ONNX Runtime (pre-built) - https://github.com/microsoft/onnxruntime/releases
-set(ONNXRUNTIME_VERSION "1.23.2")
+set(ONNXRUNTIME_VERSION "1.23.2" CACHE STRING "ONNX Runtime version")
 
 if(CMAKE_SYSTEM_NAME STREQUAL "Linux")
     if(CMAKE_SYSTEM_PROCESSOR MATCHES "x86_64|amd64|AMD64")

Dockerfile

@@ -0,0 +1,81 @@
+ARG PYTHON_IMAGE=python:3.12-slim-bookworm
+ARG POCKET_TTS_REF=058886528d0b6f2f2d4022de2e244a5260729e6e
+ARG EXPORTER_ONNXRUNTIME_VERSION=1.23.2
+ARG ONNXRUNTIME_VERSION=1.23.2
+
+FROM ${PYTHON_IMAGE} AS exporter
+ARG POCKET_TTS_REF
+ARG EXPORTER_ONNXRUNTIME_VERSION
+
+ENV PYTHONDONTWRITEBYTECODE=1 \
+    PYTHONUNBUFFERED=1 \
+    PIP_NO_CACHE_DIR=1
+
+WORKDIR /src
+
+RUN apt-get update && \
+    apt-get install -y --no-install-recommends git ca-certificates && \
+    rm -rf /var/lib/apt/lists/*
+
+COPY export_onnx.py ./
+
+RUN python -m pip install --upgrade pip && \
+    python -m pip install --index-url https://download.pytorch.org/whl/cpu torch && \
+    python -m pip install \
+        "pocket-tts @ git+https://github.com/kyutai-labs/pocket-tts.git@${POCKET_TTS_REF}" \
+        onnx \
+        "onnxruntime==${EXPORTER_ONNXRUNTIME_VERSION}" && \
+    python -m pip freeze | tee /opt/pocket-tts-exporter-requirements.txt
+
+RUN mkdir -p /opt/pocket-tts/models && \
+    python export_onnx.py --output-dir /opt/pocket-tts/models --language english_2026-01 --no-validate
+
+
+FROM ${PYTHON_IMAGE} AS builder
+ARG ONNXRUNTIME_VERSION
+
+ENV PIP_NO_CACHE_DIR=1
+
+WORKDIR /src
+
+RUN apt-get update && \
+    apt-get install -y --no-install-recommends \
+        build-essential \
+        ca-certificates \
+        git && \
+    rm -rf /var/lib/apt/lists/*
+
+RUN python -m pip install --upgrade pip cmake
+
+COPY CMakeLists.txt pocket_tts.cpp ./
+
+RUN cmake -S . -B /tmp/build -DCMAKE_BUILD_TYPE=Release -DBUILD_SHARED_LIB=OFF -DONNXRUNTIME_VERSION="${ONNXRUNTIME_VERSION}" && \
+    cmake --build /tmp/build -j"$(nproc)"
+
+RUN mkdir -p /opt/pocket-tts/runtime && \
+    install -Dm755 /src/pocket-tts /opt/pocket-tts/runtime/pocket-tts && \
+    cp -a /tmp/build/_deps/onnxruntime-src/lib/. /opt/pocket-tts/runtime/
+
+
+FROM ${PYTHON_IMAGE} AS runtime
+
+RUN apt-get update && \
+    apt-get install -y --no-install-recommends \
+        ca-certificates \
+        libgomp1 && \
+    rm -rf /var/lib/apt/lists/*
+
+WORKDIR /app
+
+ENV LD_LIBRARY_PATH=/app
+
+COPY --from=builder /opt/pocket-tts/runtime/ /app/
+COPY --from=exporter /opt/pocket-tts/models/ /app/models/
+COPY --from=exporter /opt/pocket-tts-exporter-requirements.txt /app/exporter-requirements.txt
+
+RUN mkdir -p /app/voices/.cache
+
+EXPOSE 8080
+
+ENTRYPOINT ["/app/pocket-tts"]
+CMD ["--server", "--port", "8080"]

README.md

@@ -121,7 +121,7 @@ The `/v1/audio/speech` endpoint is compatible with the OpenAI TTS API. Any clien
 ```bash
 curl -X POST http://localhost:8080/v1/audio/speech \
   -H "Content-Type: application/json" \
-  -d '{"model": "tts-1", "input": "Hello world!", "voice": "voice", "response_format": "wav"}' \
+  -d '{"model": "tts-1", "input": "Hello world!", "voice": "voice.wav", "response_format": "wav"}' \
   --output speech.wav
 ```
 
@@ -145,6 +145,27 @@ int    ptt_stream_read(void* stream_ctx, float** out_samples, int* out_len);
 void   ptt_stream_end(void* stream_ctx);
 ```
 
+### Docker
+
+This will build and run `pocket-tts-cpp:latest`:
+
+```
+docker compose up --build
+```
+
+After build can also be used like this:
+```
+docker run --rm -it -v ./voices:/app/voices -p 8080:8080 pocket-tts-cpp --server --port 8080
+```
+
+Usage same as before:
+```
+curl -X POST http://localhost:8080/v1/audio/speech \
+  -H "Content-Type: application/json" \
+  -d '{"model": "tts-1", "input": "Hello world!", "voice": "voice.wav", "response_format": "wav"}' \
+  --output speech.wav
+```
+
 ## Caching
 
 PocketTTS uses two layers of disk caching, both stored under `voices/.cache/`:

docker-compose.yml

@@ -0,0 +1,15 @@
+services:
+  pocket-tts-cpp:
+    image: pocket-tts-cpp:latest
+    container_name: pocket-tts-cpp
+    build:
+      context: .
+      args:
+        POCKET_TTS_REF: ${POCKET_TTS_REF:-058886528d0b6f2f2d4022de2e244a5260729e6e}
+        EXPORTER_ONNXRUNTIME_VERSION: ${POCKET_TTS_EXPORTER_ONNXRUNTIME_VERSION:-1.23.2}
+        ONNXRUNTIME_VERSION: ${POCKET_TTS_ONNXRUNTIME_VERSION:-1.23.2}
+    volumes:
+      - ./voices:/app/voices
+    ports:
+      - "8080:8080"
+    command: ["--server", "--port", "8080", "--precision", "${POCKET_TTS_PRECISION:-fp32}"]

export_onnx.py

@@ -1198,8 +1198,10 @@ def main():
                         help="Which models to export")
     parser.add_argument("--max-seq", type=int, default=MAX_SEQ_LEN,
                         help=f"Max KV cache length for flow_lm (default: {MAX_SEQ_LEN})")
+    parser.add_argument("--language", default="english_2026-01",
+                        help="Built-in PocketTTS config to use when --config is not set")
     parser.add_argument("--config", default=None,
-                        help="Path to config YAML (default: use built-in b6369a24)")
+                        help="Path to config YAML (overrides --language)")
     parser.add_argument("--validate-only", action="store_true",
                         help="Skip export/quantize, only validate existing ONNX files")
     parser.add_argument("--no-validate", action="store_true",
@@ -1247,13 +1249,16 @@ def _download(url: str, dest: Path):
         _download(f"{HF_BASE}/tokenizer.model", tokenizer_file)
         print(f"  ✓ {tokenizer_file}")
 
-    # Build config pointing at local weights
+    # Build config pointing at local weights/tokenizer. Upstream renamed the
+    # original b6369a24 English config to english_2026-01 in pocket-tts v2.x.
     import pocket_tts as _ptt
-    config_path = Path(_ptt.__file__).parent / "config" / "b6369a24.yaml"
     if args.config:
         config_path = Path(args.config)
+    else:
+        config_path = Path(_ptt.__file__).parent / "config" / f"{args.language}.yaml"
     config = load_config(config_path)
     config.weights_path = str(weights_file)
+    config.flow_lm.lookup_table.tokenizer_path = str(tokenizer_file)
 
     # Load model with voice cloning weights
     print("Loading model...")