sherpa-onnx

What problem does it solve?

Enable deterministic, private, and low-latency speech processing locally without internet access by providing ready guidance to run ASR, TTS, VAD, speaker diarization, speaker ID/verification, speech enhancement, audio tagging, keyword spotting, and source separation using ONNX models and the sherpa-onnx runtime.

Core Features & Use Cases

• Streaming & Non‑streaming ASR: real-time microphone transcription via OnlineRecognizer and batch/file transcription via OfflineRecognizer.
• TTS Engines: generate speech locally with Kokoro, Piper, Matcha, VITS, or KittenTTS for multi‑speaker and multi‑language needs.
• VAD, Diarization & Speaker Tasks: voice activity detection, segmentation, embedding extraction, identification and verification workflows for meetings and call analytics.
• Enhancement & Tagging: denoise or separate sources, classify audio content, and detect keywords on-device for privacy-sensitive or edge applications.
• Use Case Example: transcribe a meeting audio file on an offline workstation, split speakers using pyannote segmentation plus embeddings, and export timestamped captions and per-speaker transcripts.

Quick Start

Download the appropriate ONNX model, install sherpa-onnx, and run a local offline transcription of meeting.wav with OfflineRecognizer.from_sense_voice to produce a timestamped transcript.