Conformer-1: a new large scale/robust speech recognition model

We're introducing Conformer-1, a state-of-the-art speech recognition model trained on 650K hours of audio data that achieves near human-level performance and robustness across a variety of data.

Our results demonstrate that Conformer-1 is more robust on real-world data than popular ASR models, making up to 43% fewer errors on noisy dataI, and achieving state-of-the-art results on a wide variety of academic and real-world datasets compared to other ASR models.

▬▬▬▬▬▬▬▬▬▬▬▬ CONNECT ▬▬▬▬▬▬▬▬▬▬▬▬

▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬

#MachineLearning #DeepLearning

[魏一傑]

Whats the benefit of using this model if its not an free and open-sourced one? I can run Whisper on my own machine without any extra cost except for electricity.

[Jingizz]

could you add language picket in the dropdown? uploaded a german video but it was gibberish, but in the documentation I saw that you guys support it

[michaelzumpano]

This is brilliant! I started following Steven Brenton’s work (Univ Washington), maybe 7 or 8 years ago, when he started developing the concepts of sparse identification of non-linear dynamics (SINDy). I think sparsity, as he developed it, is going to make almost everything we do more efficient. I think it will also give us insight into AGI, because our memory and cognition both depend on sparse attention.

[danielmonge]

How does it compare with OpenAI's Whisper? Whisper was trained on 30k more hours than this model.

[AlexFord]

I really wanted to see the same audio with competitor models lol. That would have been a great juxtaposition.

[yosolonopuedo]

Nice to know that it can recognize Lebron's speech!!!!

[basrurkrishna]

In the playground, can you add a module where we can listen to the speech devoid of noise, that was used to transcribe accurately.

[lazy_iitian]

What's the pricing for conformer 1 ?

[kipchickensout]

So I tried it with a German youtube video and it got almost everything wrong in the transcription, so unless you speak as clearly as possible, for example in a podcast, and go slow, you will have a hard time

all AI models for some reason specialize in american english with a bit of other languages in between

[nivasgopi]

Hello Misra, I have a question. Generally, how are these giant models trained? I mean it requires a lot of GPUs right?? How do they manage the infrastructure to train these type of giants??

[guepardiez]

Awesome! However, I miss an option to export to a subtitle file in the playground. Any plans to implement this?