[FEAT] sherpa-onnx.rn: expose offline CT-Transformer punctuation engine via PunctuationService

[brainlybai]

Environment

• expo-audio-studio version: n/a (this is a @siteed/sherpa-onnx.rn feature request)
• Package: @siteed/sherpa-onnx.rn 1.3.1-beta.1
• Expo SDK version: 55
• Platform & OS version: iOS 17 + Android 14 (both affected)
• Device: any

Note: filed against this repo because sherpa-onnx.rn lives in this monorepo. Not related to expo-audio-studio.

Description

The package already vendors the offline CT-Transformer punctuation bindings on both platforms:

• iOS: SherpaOnnxCreateOfflinePunctuation, SherpaOfflinePunctuationAddPunct, SherpaOnnxDestroyOfflinePunctuation (C API in CSherpaOnnx)
• Android: com.k2fsa.sherpa.onnx.OfflinePunctuation (Kotlin)

But the RN PunctuationService / PunctuationHandler only reaches the online CNN-BiLSTM engine. There is no way for a JS caller to use the offline CT-Transformer models from JS today, even though the native code paths are compiled in and shipped in every install.

This blocks using the sherpa-onnx-punct-ct-transformer-zh-en-vocab272727 family of models, which produce Chinese + English punctuation in a single ~70 MB int8 model — useful for any ASR pipeline that needs both languages from one punctuator.

Expected Behavior

A JS caller can opt into the offline engine by passing a model filename in PunctuationModelConfig, e.g.:

await PunctuationService.init({
  modelDir: '/abs/path/to/punct-ct-transformer-zh-en',
  model: 'model.int8.onnx',
});
const out = await PunctuationService.addPunctuation('你好 hello world how are you');
// → '你好, hello world. How are you?'

When model is set, the native side uses the offline CT-Transformer engine and ignores cnnBilstm / bpeVocab. When model is unset, behavior is unchanged (existing online callers see no regression).

Actual Behavior

PunctuationService.init only accepts cnnBilstm + bpeVocab and always constructs an OnlinePunctuation. The offline OfflinePunctuation class is unreachable from JS.

Logs

n/a (missing feature, not a crash)

Configuration

// Today: only online engine reachable
await PunctuationService.init({
  modelDir,
  cnnBilstm: 'model.onnx',
  bpeVocab: 'bpe.vocab',
});

// Desired: offline CT-Transformer engine
await PunctuationService.init({
  modelDir,
  model: 'model.int8.onnx',
});

Fix

Wire the existing offline bindings through PunctuationHandler (Kotlin + Swift) and PunctuationService (TS). Backwards-compatible — a new optional model field opts into the offline engine; absence of model keeps the current online behavior.

PR: #410

Checklist

• Verified offline bindings already exist in iOS CSherpaOnnx and Android Kotlin
• Implemented end-to-end and validated in a downstream RN 0.85 / New Architecture app
• Existing online callers remain unchanged (verified via unit test)

[abretonc7s]

Implemented by #418 and released in @siteed/sherpa-onnx.rn@1.3.1. Closing.