David Dale

Hi! This question has already been addressed in https://github.com/facebookresearch/seamless_communication/issues/109 https://github.com/facebookresearch/seamless_communication/issues/32, and the short answer "it's complicated". For adding new languages to text translation models, there are some existing pointers, including:...

Hi! One potential solution would be the following: 1. Detect the silence and voice in the source audio using some external voice activity detection model. 2. Split the source audio...

The Seamless project did not release a speech language identification model. However, you can use a speech LID model from a related project called MMS: https://github.com/facebookresearch/fairseq/blob/main/examples/mms/README.md#lid. In https://github.com/facebookresearch/seamless_communication/issues/325 I give...

@ndlongvn If you are using the HuggingFace version of the weights, then you can refer to [the documentation of the HF re-implementation of Seamless ](https://huggingface.co/docs/transformers/v4.36.1/en/model_doc/seamless_m4t#usage)for the inference recipes. You will...

No, spoken language identification is not integrated in Seamless models. If you want it, you'll have to apply some external LID model, such as MMS. I provide more details in...

Hi @asasas234! Could you please provide some more details? * The exact model and code that you used to generate the translation; * The translation that you got as a...

The [SeamlessExpressive](https://github.com/facebookresearch/seamless_communication/blob/main/docs/expressive/README.md) model released in November does preserve many properties of the source voice. I guess, it fulfills what this issue is asking about.

The current Seamless family supports Sinhala only with the SeamlessM4T-Medium model, and only in the text modality (see https://github.com/facebookresearch/seamless_communication/tree/main/docs/m4t for more details, and model cards https://github.com/facebookresearch/seamless_communication/tree/main/src/seamless_communication/cards for the lists of...

It is the same issue as https://github.com/facebookresearch/seamless_communication/issues/174. The solution is to translate one sentence at a time, because this is how the model has been trained.

Actually, both Seamless and its text-only predecessor NLLB were trained mostly on single-sentence translation, and they are by no means guaranteed to correctly translate multiple-sentence texts. Thus, the safest recommendation...