Multi language support
Are there any plans for a version of this model supporting spanish and portuguese languages ?
I hope that multilingual support will be implemented. This would be a great second-layer solution for customer support on my sales website. This AI is excellent at dealing with people!
Please include Portuguese.
Include french too please
Feedback taken! Multilingual is hard because 1) Moshi is not multilingual so multilingual finetunes will be quite limited so swapping base models needs to be figured out first 2) the edge that this model has over standard ASR+LLM+TTS voice agent stack is naturalness that comes from real channel-separated conversations dataset, where Fisher English is the only publicly available source.
Will keep this discussion open to hear more language requests, and pointers to publicly available, commercially-available channel-separated dialog datasets in other languages.
is italian possible perhaps?
is Bangla possible? please Add Bangla language support.
French support would also be great
Feedback taken! Multilingual is hard because 1) Moshi is not multilingual so multilingual finetunes will be quite limited so swapping base models needs to be figured out first 2) the edge that this model has over standard ASR+LLM+TTS voice agent stack is naturalness that comes from real channel-separated conversations dataset, where Fisher English is the only publicly available source.
Will keep this discussion open to hear more language requests, and pointers to publicly available, commercially-available channel-separated dialog datasets in other languages.
Hi Royra, how many hours of channel-separated-conversation are the minimum required to get a good finetuning for a new language ?
the edge that this model has over standard ASR+LLM+TTS voice agent stack is naturalness that comes from real channel-separated conversations dataset, where Fisher English is the only publicly available source.
Will keep this discussion open to hear more language requests, and pointers to publicly available, commercially-available channel-separated dialog datasets in other languages.
Hi @royrajarshi , I’m an engineer at oto. We have Japanese and English channel-separated conversational speech datasets. We also run a dedicated conversation data-collection platform and would love to expand to more languages. We’ve released English datasets publicly (e.g., https://huggingface.co/datasets/otoearth/otoSpeech-full-duplex-processed-141h), and we also have additional datasets available under a commercial license.
We’d love to collaborate and can share more samples if helpful. Happy to chat.