Ning comments

Results 7 comments of


                                            Ning

Cannot import 'latent_var_criterion'

Will land a fix soon. You can comment out latent_var related imports in train.py now as a workaround.

beam search

Sorry README is out-of-date. We already have BeamSearch class fully scripted in ensemble_export.py. Also Pytorch->ONNX->Caffe2 export path as mentioned in README is not supported now. We are switching to TorchScript...

Is it possible to generate subtitles like whisper ai?

Hey @barinov274 - It's not trivial to get ASR timestamps for our model unfortunately. Since it shares with translation tasks, decoding process is not "monotonic" like other ASR approaches (e.g....

Poor transcription in less than 60 sec audio

Hey @Azam2107 - We suggest cutting the audio to

Poor transcription in less than 60 sec audio

For English ASR that could be a general quality issue we are investigating.

Low quality results compared to Huggingface spaces and Demo!

Could you compare your implementation with HF demo around https://huggingface.co/spaces/facebook/seamless_m4t/blob/main/app.py#L78? Note that resampling might be needed.

[WIP, don't merge] unity.cpp -> ggml master

Thanks @ggerganov for detailed comments! Regarding 1 and 2, makes sense, I will keep unity.cpp under seamless_communication for now. So keeping KNF library there would be easier without bloating ggml...