Add language hint for Omnilingual ASR CTC on Android & iOS

[anita-smith1]

Many thanks for adding Omnilingual ASR CTC support to sherpa-onnx.

I’m using the sherpa-onnx-omnilingual-asr-1600-languages-300M-ctc-int8-2025-11-12 model on Android (type 44 in getOfflineModelConfig). The Android bindings expose:

data class OfflineOmnilingualAsrCtcModelConfig(
    var model: String = "",
)

In many mobile use cases, however, the target language is already known (e.g. “always English” or a fixed eng_Latn / cmn_Hans etc.). It would be very useful to be able to pass a language hint from Android to bias or restrict decoding to that language.

Feature request

Add an optional language field to the Android config, for example:

data class OfflineOmnilingualAsrCtcModelConfig(
    var model: String = "",
    var language: String = "", // optional language hint, e.g. "eng_Latn"
)

[csukuangfj]

sorry, this model does not support that.

Note that we are using the 300M CTC ASR of the Ominlingual model.

There are many variants. I hope you know the differences among.them.

[anita-smith1]

@csukuangfj Okay thank you. Is there any way to improve accuracy by restricting to a particular language. Any way at all on the application layer and not the model? Also is it possible to finetune or train an omnilingual model for just one language and convert to sherpa onnx? sorry if my questions is bad

[csukuangfj]

Is there any way to improve accuracy by restricting to a particular language.

Sorry, I don't think there is such a way for this specific model. If you use their LLM variant, it may help; but we don't support it.

---

Any way at all on the application layer and not the model?

Sorry, no.

---

Also is it possible to finetune or train an omnilingual model for just one language

Please ask at https://github.com/facebookresearch/omnilingual-asr/issues

and convert to sherpa onnx

Yes, i think it is doable.