sherpa-onnx

The SenseVoice Nano models in this directory are converted versions of the Fun-ASR-Nano-2512 models from HuggingFace. These models are optimized for high-performance speech recognition in challenging environments and diverse linguistic scenarios.

Key Capabilities:

• Far-field & High-noise Recognition: Optimized for far-distance pickup and noisy environments (e.g., conference rooms, vehicles, industrial sites) with up to 93% accuracy.
• Chinese Dialects & Accents: Supports 7 major dialects (Wu, Cantonese, Min, Hakka, Gan, Xiang, Jin) and covers 26 regional accents (including Henan, Shaanxi, Hubei, Sichuan, Chongqing, Yunnan, Guizhou, Guangdong, Guangxi, etc.).
• Multi-language Support: Supports 31 languages, with specific optimization for East and Southeast Asian languages. It supports free language switching and mixed-language recognition.
• Music Background Recognition: Specifically enhanced to recognize lyrics accurately even when music is playing in the background.

This directory provides several implementation examples for performing streaming Automatic Speech Recognition (ASR) using the Dart API. Each example is tailored to a specific model architecture. To run these examples, use the provided shell scripts which handle the necessary environment setup and execution parameters.

| File | Description|
|------|------------|
|[./bin/nemo-transducer.dart](./bin/nemo-transducer.dart)| Use a NeMo transducer model for speech recognition. See [./run-nemo-transducer.sh](./run-nemo-transducer.sh)|
|[./bin/paraformer.dart](./bin/paraformer.dart)| Use a Paraformer model for speech recognition. See [./run-paraformer.sh](./run-paraformer.sh)|
|[./bin/zipformer-ctc-hlg.dart](./bin/zipformer-ctc-hlg.dart)| Use a Zipformer CTC model with HLG graph for speech recognition. See [./run-zipformer-ctc-hlg.sh](./run-zipformer-ctc-hlg.sh)|
|[./bin/zipformer-ctc.dart](./bin/zipformer-ctc.dart)| Use a Zipformer CTC model for speech recognition. See [./run-zipformer-ctc.sh](./run-zipformer-ctc.sh)|
|[./bin/zipformer-transducer.dart](./bin/zipformer-transducer.dart)| Use a Zipformer transducer model for speech recognition. See [./run-zipformer-transducer.sh](./run-zipformer-transducer.sh)|

The dotnet-examples directory provides various C# implementations for using sherpa-onnx features. Key examples include:

• Speech Enhancement (Offline):
  • speech-enhancement-gtcrn: Uses GTCRN models for offline speech denoising.
  • speech-enhancement-dpdfnet: Uses DPDFNet models for offline speech denoising. For downstream ASR, use 16 kHz models like dpdfnet_baseline.onnx, dpdfnet2.onnx, dpdfnet4.onnx, or dpdfnet8.onnx. For 48 kHz enhancement output, use dpdfnet2_48khz_hr.onnx.
• Speech Enhancement (Online/Streaming):
  • streaming-speech-enhancement-gtcrn: Uses GTCRN models for online speech denoising.
  • streaming-speech-enhancement-dpdfnet: Uses DPDFNet models for online speech denoising.
• Text-to-Speech (TTS):
  • zipvoice-tts: Uses ZipVoice for Chinese/English zero-shot text-to-speech.
  • zipvoice-tts-play: Uses ZipVoice for Chinese/English zero-shot text-to-speech with audio playback.

For full API details, refer to the official C# API documentation.

The harmony-os directory contains several implementation examples for running speech and audio tasks on-device:

• Speaker Diarization: Implementation of on-device speaker diarization (SherpaOnnxSpeakerDiarization).
• Speaker Identification: Using speaker embedding models for on-device identification (SherpaOnnxSpeakerIdentification).
• Streaming ASR: Real-time on-device speech recognition using streaming ASR models (SherpaOnnxStreamingAsr).
• Text-to-Speech (TTS): Running on-device TTS (SherpaOnnxTts). See documentation at https://k2-fsa.github.io/sherpa/onnx/harmony-os/tts.html.
• VAD + Non-streaming ASR: Combining Voice Activity Detection (VAD) with non-streaming ASR for speech recognition (SherpaOnnxVadAsr). See documentation at https://k2-fsa.github.io/sherpa/onnx/harmony-os/vad-asr.html.

This directory provides several Dart-based examples for performing non-streaming Automatic Speech Recognition (ASR). Each example is tailored to a specific model architecture or model provider. To run these examples, use the corresponding .sh shell script provided for each binary.

| File | Description |
|------|------------|
|[./bin/dolphin-ctc.dart](./bin/dolphin-ctc.dart)| Use a Dolphin Ctc model for speech recognition. See [./run-dolphin-ctc.sh](./run-dolphin-ctc.sh)|
|[./bin/nemo-ctc.dart](./bin/nemo-ctc.dart)| Use a NeMo Ctc model for speech recognition. See [./run-nemo-ctc.sh](./run-nemo-ctc.sh)|
|[./bin/nemo-transducer.dart](./bin/nemo-transducer.dart)| Use a NeMo transducer model for speech recognition. See [./run-nemo-transducer.sh](./run-nemo-transducer.sh)|
|[./bin/paraformer.dart](./bin/paraformer.dart)|Use a paraformer model for speech recognition. See [./run-paraformer.sh](./run-paraformer.sh)|
|[./bin/telespeech-ctc.dart](./bin/telespeech-ctc.dart)| Use models from Tele-AI/TeleSpeech-ASR for speech recognition. See [./run-telespeech-ctc.sh](./run-telespeech-ctc.sh)|
|[./bin/whisper.dart](./bin/whisper.dart)| Use whisper for speech recognition. See [./run-whisper.sh](./run-whisper.sh)|
|[./bin/zipformer-transducer.dart](./bin/zipformer-transducer.dart)| Use a zipformer transducer for speech recognition. See [./run-zipformer-transducer.sh](./run-zipformer-transducer.sh)|
|[./bin/vad-with-paraformer.dart](./bin/vad-with-paraformer.dart)| Use a silero-vad with paraformer for speech recognition. See [./run-vad-with-paraformer.sh](./run-vad-with-paraformer.sh)|
|[./bin/sense-voice.dart](./bin/sense-voice.dart)| Use a SenseVoice CTC model for speech recognition. See [./run-sense-voice.sh](./run-sense-voice.sh)|