sherpa-onnx for local speech applications

How it works

The project supplies APIs and examples for speech tasks such as recognition and synthesis on different platforms. Streaming and offline workflows use different model and configuration choices.

What you need

Use a model and package combination supported on the target platform. Keep tokenizer or token files and streaming configuration with the model artifact.

What to try first

Try a documented streaming recognizer on the intended device. Measure partial-result latency and verify final text when the user pauses mid-sentence.

Measure the delay between end of speech and a final transcript. A recognizer that emits fast partial text can still take too long to close an utterance.

Where to look for failures

Each model has its own expected sample rate, token files and decoding settings. Verify the exact platform package and model pair.

Related Franklin pages

Test startup, partial results and end-of-utterance delay on the intended device. An offline throughput figure does not capture how long a user waits to see the first text.

• Add local speech-to-text to applications through a language-specific model and API. Vosk for offline streaming speech recognition
• Generate speech locally using published Kokoro voices and inference tooling. Kokoro for compact text-to-speech pipelines

For an application where the quality of recording and handling audio matters, see Franklin’s Open Mic AI profile. Open Mic AI

Project documentation

• Current project README
• README at the reviewed revision