Cohere's new AI model can transcribe your audio
Cohere just launched Transcribe, its first open-source tool for automatic speech recognition.
It's lightweight enough to run on regular GPUs and supports 14 languages like English, French, and Chinese.
Made for things like note-taking and analyzing audio, it's intended to be accessible to users who can self-host on consumer-grade GPUs.
Transcribe outperforms popular models on Hugging Face Open ASR leaderboard
Transcribe outperforms several popular models on the Hugging Face Open ASR leaderboard, achieving a 5.42 WER and a reported 61% average win rate in human evaluations and can process a massive 525 minutes of audio in just one minute.
While it still has some trouble with Portuguese, German, and Spanish, Cohere reports that human evaluators gave Transcribe a 61% average win rate over other models.
Cohere plans to roll it out across its enterprise platform North and make it available through APIs, so expect to see this tech pop up in more apps soon.