How to specify language and region in Dolphin ASR to avoid misrecognition?

[ctop007]

Hello, I’m using Dolphin for ASR recognition and would like to know if it is possible to specify the language and region during inference.

In my testing, Dolphin performs significantly better than Whisper in terms of recognition quality. However, when I don’t specify the language or region, the recognition results often become confusing. For example, Arabic speech is sometimes incorrectly recognized as Japanese.

Could you please let me know how I can explicitly set the language and region for recognition?

Thanks!

[csukuangfj]

Sorry, not supported in sherpa-onnx.

Please see #2293 for more details.

[ctop007]

That’s really unfortunate.

By the way, do you know if there are any ASR models that work better for Arabic? I tried Whisper with explicit language specification, but the recognition quality is still not very good.

Also, would it be possible to improve accuracy by using a community fine-tuned single-language version of Whisper? Currently, I only see Chinese and English models available. Can our system support extending to other languages (e.g., Arabic) through such fine-tuning?

Thanks again!

[csukuangfj]

Please try
https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/sherpa-onnx-streaming-zipformer-ar_en_id_ja_ru_th_vi_zh-2025-02-10.tar.bz2
with sherpa-onnx

[csukuangfj]

Also, you can convert
https://huggingface.co/MostafaAhmed98/Conformer-CTC-Arabic-ASR/tree/main
to sherpa-onnx

[csukuangfj]

I suggest that you use google to look for models for arabic asr.

[csukuangfj]

Also, this one

https://huggingface.co/nvidia/stt_ar_fastconformer_hybrid_large_pcd_v1.0

there are a lot.

[ctop007]

Thank you so much for sharing these model links! I’ll go through them one by one and test their performance.

By the way, regarding the model you mentioned here:

Please try https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/sherpa-onnx-streaming-zipformer-ar_en_id_ja_ru_th_vi_zh-2025-02-10.tar.bz2 with sherpa-onnx

Does this one support explicit language specification during recognition (e.g., forcing it to recognize Arabic only)?

Thanks again for the detailed guidance! 🙏

[csukuangfj]

no. I don't think you can find such an open-source ASR model.