support for ndarray and tensor decoding for RNNT based ASR models about nemo

[titu1994]

It's already in the main branch, you can try it out there. However, the next release will be probably at the end of the month. But it will probably be a prerelease

[titu1994]

No, it will be the next major version release

[gordon0414]

@titu1994

[titu1994]

The docstring isn't updated, but yes main branch supports rnnt as of right now - https://github.com/NVIDIA/NeMo/blob/main/nemo/collections/asr/models/rnnt_models.py#L274

[gordon0414]

The docstring isn't updated, but yes main branch supports rnnt as of right now - https://github.com/NVIDIA/NeMo/blob/main/nemo/collections/asr/models/rnnt_models.py#L274

Thanks for the quick reply, but if we check the input arguments right now,
https://github.com/NVIDIA/NeMo/blob/main/nemo/collections/asr/models/rnnt_models.py#L237

   @torch.no_grad()
    def transcribe(
        self,
        audio: List[str],
        batch_size: int = 4,
        return_hypotheses: bool = False,
        partial_hypothesis: Optional[List['Hypothesis']] = None,
        num_workers: int = 0,
        channel_selector: Optional[ChannelSelectorType] = None,
        augmentor: DictConfig = None,
        verbose: bool = True,
        override_config: Optional[TranscribeConfig] = None,
    ) -> TranscriptionReturnType:

the audio is only limited to a list of strings, in which this case a path to the audio files.

[titu1994]

That's a type annotation, it has no effect on the actual execution of code. That type annotation is wrong. But the code is going to execute RNNT with tensor just fine cause the underling model supports it

[gordon0414]

thanks for the quick reply again, maybe I should try again!

[gordon0414]

One last question, is the Code Refactoring for ASR applied on 1.23.0 ??

[gordon0414]

any estimation of release date yet??

[gordon0414]

thx, I tried with main branch and it works smooth like butter!
keep up the great work!!!