[Help needed] Conformer-Transducer Streaming ASR with Microphone

[ItsaiNk]

Hi, I really need help. I trained a Conformer-Transducer model and I have to online streaming inference with the audio recorded by a microphone, like Online_ASR_Microphone_Demo.ipynb. I also read the tutorials Streaming_ASR.ipynb and Buffered_Transducer_Inference.ipynb but I'm not understanding how to combine all this example to performe what I need.
Someone can help me please? Thanks!

[titu1994]

Microphone support is quite fragile, and those scripts cannot easily be translated to it.

If you need it just for a demo, a very simple approach would be to use a gradio app in streaming mode.

For example, you can extend the code here as needed - https://huggingface.co/spaces/smajumdar/nemo_conformer_rnnt_large_streaming

[ItsaiNk]

Thank you for the answer! There is a way to pass the audio stream and not save a wav file before trascription?

[titu1994]

Currently, there isn't support for that. But we're eventually designing a more seperable transcription API that can take tensors as inputs.