Paper: Repo and Dataset:

Really good comments here: in French mainly, but you’ll discover quite the interesting companies! :)

@GuillaumeRozier @Vocca_ai 😇

A big big interview tonight from Emmanuel Macron, the French President, regarding AI! What he said: 🤖 Models should be considered as assistants, not replacements: AI will be able to reintroduce humans by lowering mind-numbing tasks 🗒️ Regulation will not be French, not European, but Worldwide as it is a worldwide technology: However, AI generated content needs to be marked as such 🎯 Europe MUST invest in AI ; regulation needs to come once Europe is in the run rather than while it is trying to get in 🌎 Social media algorithms must be transparent, and the regulation effort should be worldwide 💸 Europe announces 109B of Euros of investment in AI to bolster its companies ; jointly with EUA, Canada, or American Firms 🍃 Regarding Data Center consumption, French nuclear+green electricity is a real strong advantage and Frugal AI along Durable Data Center construction is in the works ⚙️ The word accelerate was pronounced multiple times: Draghi doctrine in action 🇮🇳 The Indian partnership is a way to strengthen Europe independence, as part of the alliance between Europe and the Indo-Pacific ; the goal is to form a 3rd Pathway against US and China as a major power 🔎 Technologies from other countries like China or the Us will not be banned, but scrutinized to watch over technological sovereignty A really really good interview, giving hope for the Future of France (and as a result, Europe) as a major actor in AI!

On the one side I think there could be real value in Finetuning reasoning models for medical diagnosis On the other side I fear it might end up with the model overthinking and telling us we have cancer for even the most benign disease Curious where it would lead

Paper GitHub

Actually disagreeing : I think it will just make consultant more powerful Sure the workforce required will be reduced, but if making researches & presentations is easier, the person which is paid to make that will just be able to make 10 presentations in a day instead of one It will increase reach & market competitiveness by giving people the tools to be more efficient, which means giants can be toppled if they overlook the wave But it’s just like translation imo: what I think is instead that 10x consultants/translators can be born out of the AI revolution, or in fact that even 1x consultants can be turned into 10x A reason you want someone external is often because of budget, politics or to offload time. You will still need a human to read the output work, appropriate it and make sure it is valid imo

@dtiapkin @ramealexandre Oh Very very curious, will read it

@maxencejm Eager to see what comes out!

@AymericRoucher Very interesting! Got the link to the benchmark introduction?

The second Mistral-R1 comes online OpenAI and Anthropic might have a competitor over B2B dev questions Right now o3-mini can answer most questions much better than everything else (especially technical precise stuff) but takes around 30s to answer which is perceivable as long If there is a very fast reasoning model with coding performances that are high, LeChat might just steal the dev market from the two other companies

@euacchq @MistralAI It’s indeed super fast All that the product needs now is fast reasoning models and it might become the goto for coding

RT @ADarmouni: Hibiki : An open-source decoder-only realtime translation model 🧵 📖 Read of the day, season 3, day 15: « High-Fidelity Simu…

Paper link right here: What makes me positive as well is that in fact their training process is not language dependent. Their training data is language dependent, but you just need to synthesize quality English -> French or Spanish -> English if you want to improve capacities. I am thus very curious if one model can take into account more languages, or if there needs to be a Hibiki for English -> French, one for Spanish -> English and so on. Hyped. ;)

📂 Hibiki is then evaluated in several metrics, and manages to challenge SoTA pretty well : - Very good ASR with High Speaker Similarity on both Short Form and Long Form audio evaluation datasets - Human Evaluation places it at a very good rating. Ablations confirm the authors’ choices.

🔧 Afterwards, Hibiki is lastly finetuned on a quality speech synthetic translation dataset (created with the trained TTS model), made of 900 hours of long form utterances and samples with natural pauses and high speaker similarity. Now, for the smaller Hibiki model: this one is going through the same audio and text pretraining steps, but is soft distilled in the speech translation part.

⚙️ A TTS model is thus specifically trained to take into input both a text and a voice to condition. Model can output both streamed text and speech. The text streamed is conditioned to match exactly the text input, and thus all the model can do is insert padding tokens to create pauses, or hasten or decelerate the generated speech. If the TTS model lags on the audio that needs to be generated, a penalty is given on the logits of the padding token, which enforces smoothness. For each of the Fr -> En transcriptions, 6 to 8 generations per input were done. Samples were in the end selected first on WER, and then Speaker Similarity (cosine similarity within embeddings of Input and Output Audios). A conditioning on 10s of the French audios was applied. A silence insertion technique is used so that the translated audios and reference audios are as close as possible in utterance. The data preparation is not over though!