speech_sherpa 0.1.0

Cross-platform speech provider over sherpa-onnx - streaming and batch STT, VAD, diarization, and speaker embeddings.

Changelog

0.1.0

• Initial package: cross-platform batch speech-to-text, Silero voice-activity detection, pyannote + WeSpeaker diarization with provenance-tagged speaker embeddings, and an on-demand model registry, all over sherpa-onnx.
• Streaming speech-to-text over sherpa's OnlineRecognizer, emitting volatile RecognitionPartial hypotheses that are confirmed as RecognitionFinal when the endpointer closes a segment. Runs on its own long-lived worker isolate.
• Streaming Zipformer models in the catalog: streamingZipformerEn (English, 296 MB, the streaming default) and streamingZipformerEn20M (English, 122 MB, low-power). Recognition capability is now declared per model, so a batch model never advertises streamingSpeechToText and vice versa.
• SherpaStreamingRecognitionOptions exposes sherpa's three endpointing rules. They are silence timers over decoder state, not an end-of-utterance model.
• transcribe refuses audio shorter than SpeechAudioGuards.minimumRecognitionDuration with the typed speech_audio_too_short failure, the same guard every batch provider enforces. This replaces the empty-audio path that returned an empty transcript: "too short to try" and "the user said nothing" are different answers and only one of them is actionable.