- Rust 96%
- Python 4%
- Shell <1%
heyo #
Real-time voice processing in Rust. Speech recognition, text-to-speech, voice activity detection and turn scoring, running locally on the CPU.
Uses Parakeet, Pocket TTS, Silero VAD and Smart Turn, with HTTP and WebSocket APIs.
Run #
Rust 1.97.1 is the tested toolchain. Supports x86-64 and AArch64; the default x86 build requires AVX2/FMA (x86-64-v3).
git lfs install
git lfs pull
cargo build --release --workspace --locked
cp heyo.example.toml heyo.toml
# Set your model and voice paths in heyo.toml.
target/release/heyo --config heyo.toml
Model weights and 19 voice recordings are included through Git LFS (about 314 MB).
The asset manifest lists file sizes, SHA-256 hashes,
licenses and original-checkpoint download links.
Models, voices and their licenses live together under assets/.
The engine binaries provide
import or calibration commands for preparing models. Configure the converted
models using heyo.example.toml,
removing sections for engines you don't need. Paths are relative to that file.
The server defaults to 127.0.0.1:8090. Set HEY_API_KEY for Bearer
authentication and use a TLS proxy when exposing it over a network.
Without that variable, the API does not require authentication.
Speech sentence budget #
[tts].max_tokens_per_chunk sets the prepared text-token budget for both REST
speech and phrase WebSocket/Live speech. It defaults to 96; operators can choose
another positive value, including 50. Whole sentences stay together when they
fit. Longer text falls back to substantial comma/semicolon/colon clauses, then
word boundaries, without dropping or reordering words.
This policy is separate from the model's context capacity. At startup, each configured voice is encoded and its actual KV positions are reserved together with the requested text tokens and the engine's worst-case generation frames. An unsafe setting fails before readiness with the voice name and computed maximum; it is never silently capped or used to truncate audio. Longer voice references leave less text capacity. The native per-generation capacity guard remains active as a final check. Model and voice changes can therefore require adjusting this setting.
API #
GET /v1/modelsand/v1/audio/voicesPOST /v1/audio/transcriptionsand/v1/audio/speech- WebSocket
/v1/audio/transcriptions/streamand/v1/audio/speech/stream
Streaming uses WebSocket subprotocol hey. Message schemas live in
crates/protocol; heyo-probe
is an example client.
Development #
cargo fmt --all --check
cargo test --workspace --locked
cargo clippy --workspace --all-targets --locked -- -D warnings
One reference test needs an external ONNX fixture and is skipped by default.
License #
AGPL-3.0-only. Third-party notices are in NOTICE. Model weights and voice recordings retain their own licenses; see asset notices.