speech

Speech Examples

Four runnable examples in one binary, selected at runtime with --usecase:

Usecase	Description
vad	VAD-driven Speech-to-Text (Silero VAD + Parakeet Nemotron)
stt	Transcribe a WAV file (Parakeet Nemotron)
tts	Synthesise speech from text (Pocket-TTS)
agent	Full voice loop: VAD → STT → LLM Agent → TTS

Prerequisites

suod apt update
sudo apt-get install -y pkg-config libasound2-dev

Set HF_TOKEN or HUGGINGFACE_HUB_TOKEN if the repository requires authentication.

Building

No feature flags are needed — all examples are always compiled:

cargo build -p speech-examples --release

Add --features cuda to enable CUDA/TensorRT acceleration for Parakeet.

Usage

VAD + STT — microphone

cargo run -p speech-examples --release -- --usecase vad --input mic --max-seconds 30

VAD + STT — WAV file

cargo run -p speech-examples --release -- --usecase vad --input file --audio-file /path/to/audio.wav

STT only — WAV file (Nemotron)

cargo run -p speech-examples --release -- --usecase stt --audio-file /path/to/audio.wav

TTS only

cargo run -p speech-examples --release -- --usecase tts --text "Hello from AutoAgents"

To download sample audio

wget https://dldata-public.s3.us-east-2.amazonaws.com/2086-149220-0033.wav -O sample.wav

Use --voice to select a Pocket-TTS voice (default: alba) and --output-file to save the audio to a WAV file.

Agent loop (VAD → STT → LLM → TTS)

export OPENAI_API_KEY=sk...
cargo run -p speech-examples --release -- --usecase agent --input mic --max-seconds 60

Flag	Default	Description
--input	mic	mic or file
--audio-file	—	WAV file path (required for --input file and stt)
--max-seconds	30	Stop mic capture after this many seconds (0 = run until Ctrl+C)
--language	—	Language hint passed to the STT model
--text	built-in	Text to synthesise (TTS only)
--output-file	—	Save TTS output to a WAV file
--agent-model	gpt-4o-mini	OpenAI model used by the agent
--voice	alba	Pocket-TTS voice identifier

Notes

• Input audio is automatically resampled to 16 kHz mono before VAD/STT processing.
• The segmenter is tuned for low-latency real-time use: 30 ms VAD windows, 450 ms silence timeout, conservative speech/silence thresholds.
• Agent mic input uses Parakeet Nemotron streaming with VAD end-of-utterance detection (English only). Ensure the nemotron-speech-streaming-en-0.6b model directory is available in the HuggingFace repo.
• The agent example requires OPENAI_API_KEY to be set.
• Say "goodbye" or "stop" to trigger the agent's exit_conversation tool and end the session cleanly.

FYI: The Parakeet ONNX models (downloaded separately from HuggingFace) by NVIDIA. This library does not distribute the models.