David Dale

In the official Seamless implementation (based on fairseq2), there is currently no recommended way to force the model to include some particular word in the output. In the Hugginface transformers...

Да, похоже на то. В целом, pickle конечно не очень надежный способ хранения моделей. Нужно будет мне поэкспериментировать с какими-то более устойчивыми, если вообще какой-то интерес к проекту имеется.

ASR with Seamless is treated as a special case of translation, where the source and target languages are the same. But the Seamless models were not trained to predict the...

Hi everyone! I was also concerned with the behavior of the NLLB tokenizer at HF, so, even before discovering this issue, I made two of my own "experiments" to verify...

Of course we can! A challenging part would be to properly align the tokens from the language model and from Seamless. I am not sure there is code that you...

Seamless doesn't support large inputs. You need to split the audio to chunks below 30 seconds before feeding it to the model.

For text inputs, Seamless indeed needs the source language code. This language code can be predicted with the NLLB text LID model (https://github.com/facebookresearch/fairseq/tree/nllb?#lid-model) which has language codes mostly consistent with...

@asusdisciple thank you for the analysis! Based on the [Glottolog database](https://glottolog.org/) and a little bit of common sense, the [M4T languages](https://github.com/facebookresearch/seamless_communication/tree/main/docs/m4t#supported-languages) that are "missing" from the [MMS language list](https://github.com/facebookresearch/fairseq/tree/main/examples/mms) can...

> However is there any way to set the target language to the source language for ASR? For example when I do not want a translation and just pure transcription...

Hi! As the Fairseq code for NLLB is not very actively supported, my recipe for adding a new language to the Huggingface implementation of NLLB might be relevant: https://cointegrated.medium.com/a37fc706b865.