@sherpaw/xsai-transcription

xsai-style streaming transcription provider for Sherpa-ONNX WASM.

This package now follows a provider/executor design similar to xsai-transformers:

• provider: createSherpawProvider().speech(...)
• execution: streamTranscription({ ...providerResult })

Structure

• src/stream-transcription/core.ts
  • low-level module/session helpers (initTranscriptionModule, createSession, etc.)
• src/stream-transcription/provider.ts
  • createSherpawProvider
  • implements fetch-override RPC adapter
• src/stream-transcription/execute.ts
  • streamTranscription(...) final API executor
• src/stream-transcription/session.ts
  • runtime streaming session built on @sherpaw/asr
• src/stream-transcription/events.ts
  • event communication via @moeru/eventa

Main API (xsai-style)

import {
  asRemoteUrl,
  createSherpawProvider,
  streamTranscription,
} from '@sherpaw/xsai-transcription'
import transcriptionWorkerURL from '@sherpaw/xsai-transcription/worker?worker&url'

const sherpawProvider = createSherpawProvider({
  workerURL: transcriptionWorkerURL,
})

const speech = sherpawProvider.speech({
  metadata: asRemoteUrl('https://huggingface.co/.../tokens.json'),
  data: asRemoteUrl('https://huggingface.co/.../model.int8.onnx.data'),
  sampleRate: 16000,
})

await speech.loadSpeech()

const transcription = streamTranscription({ ...speech })
const writer = transcription.input.getWriter()
await writer.write(float32Chunk)
await writer.close()
const final = await transcription.done

const sentenceReader = transcription.streams.sentences.getReader()
const { value: sentence } = await sentenceReader.read()
console.log('sentence:', sentence?.text)

Provider design

createSherpawProvider() returns:

• speech(model) -> { baseURL, fetch, loadSpeech, terminateSpeech }
• speech(model) exposes openStream(...) for native bi-directional command/event streaming.

streamTranscription now uses the duplex stream transport (openStream) as the primary execution path.

loadSpeech() is optional preloading for model assets before streaming starts. terminateSpeech() terminates provider-owned worker state for the speech transport instance.

Use asRemoteUrl(...) when model assets should be fetched on the main thread before being transferred into the worker.

Event model

Events emitted in streamTranscription(...).streams.full:

• transcription.started
• sentence.begin
• transcription.partial
• word
• sentence.end
• transcription.completed

These are routed with @moeru/eventa internally.

streamTranscription(...).streams also exposes typed stream branches:

• full: all events (TranscriptionEvent)
• partials: only transcription.partial
• words: only word
• sentences: only sentence.end

Low-level APIs

You can still use lower-level primitives directly:

• initTranscriptionModule()
• loadModelFiles(module, { metadata, data })
• createStreamingTranscriptionSession(...)
• createSession(module, options)
• transcribeOnce(...)
• pcm16ToFloat32(...)

Worker entry

@sherpaw/xsai-transcription/worker is still available for worker-based usage.