soundstream-light 0.1.4

No project description provided

soundstream-light

Minimal SoundStream encoder/decoder tooling with a native C++ CLI and a Python API.

Python Usage (PyPI)

Install package with uv,

uv pip install soundstream-light

Use soundstream_light in Python,

from scipy.io import wavfile
import numpy as np
from soundstream_light import encode, decode

rate, pcm = wavfile.read("test.wav")
waveform = pcm.astype(np.float32) / 32768.0

embeddings, meta = encode(waveform)
recon = decode(embeddings, metadata=meta)
wavfile.write("test_recon.wav", rate, (np.clip(recon, -1.0, 1.0) * 32768.0).astype(np.int16))

The module auto-downloads models on first use if they are missing.

Fetch Models

Use the bundled CLI to download the required .tflite files manually (defaults to ./models or SOUNDSTREAM_MODELS_DIR). If you are working from the repository, prefix the command with uv run so the local package is used.

uv run soundstream-cli models fetch

With models present, you can test the encoder/decoder round trip using the bundled test.wav example:

uv run soundstream-cli encode test.wav test.embeddings
uv run soundstream-cli decode test.embeddings test_recon.wav

Evaluate Round-Trip Quality

Install optional metric dependencies (STOI/PESQ) when you need quality numbers:

uv pip install 'soundstream-light[metrics]'

With models present, you can run the built-in evaluator against the bundled test.wav (or any mono 16 kHz WAV) to get SI-SNR, STOI, optional PESQ, and RTF:

uv run soundstream-cli evaluate test.wav --compute-pesq --save-recon test_eval.wav

Sample results on the bundled test.wav (M1, XNNPACK, 1 thread):

SI-SNR: 7.59 dB
STOI:   0.9676
PESQ:   3.6041
RTF:    0.015
Encode: 0.017 s | Decode: 0.023 s

Build from source

You can also build from source code by cloning this repo. Make sure that git, cmake, a C++17 compiler, and typical build tools (e.g. make, ninja, perl) are available.

Prepare TensorFlow Lite headers and libraries:

bash scripts/setup_tflite_from_source.sh --clean --branch v2.15.0 --generator Ninja --build-type Release

Configure and build:

cmake -S . -B build \
  -DTFLITE_SOURCE_DIR="$(pwd)/third_party/tflite-src" \
  -DTFLITE_BUILD_DIR="$(pwd)/third_party/tflite-src/cmake_build"
cmake --build build --target soundstream_cli

The binary appears at build/soundstream_cli. Then build PyPI artifacts from source,

export TFLITE_SOURCE_DIR="$(pwd)/third_party/tflite-src"
export TFLITE_BUILD_DIR="$(pwd)/third_party/tflite-src/cmake_build"
uv build

Verify the Build

Run the C++ round-trip test (requires models):

# Download models once if you haven't already
uv pip install dist/soundstream_light-*.whl
uv run soundstream-cli models fetch

cmake --build build --target soundstream_roundtrip_test
ctest --test-dir build --output-on-failure

Python unit tests can be executed without compiling the extension:

uv run python -m unittest tests.test_api

License

Apache License 2.0 — see LICENSE.

Acknowledgements

This project builds on the source code of Lyra.