In March, Transcribe topped the OpenASR leaderboard for general-purpose speech recognition. Today, it leads a benchmark designed to go beyond and test robustness in real-world, far-field audio environments.

Give it a try and share back what you build: https://huggingface.co/CohereLabs/cohere-transcribe-03-2026

Cohere Transcribe, our open-source speech recognition model, is #1 on the new @huggingface Far-Field ASR benchmark.

These tests measure performance in varying signal-to-noise conditions: the kinds of audio found in meeting rooms, contact centres, & phone calls.

In other words, environments where enterprise speech applications actually operate. Cohere Transcribe ranked #1 across every metric:

Transcribe achieved a 17.9 WER - nearly 2 points ahead of IBM Granite Speech and 3.6 points ahead of NVIDIA’s Parakeet.

Still Apache 2.0 and runs on your laptop. Enterprise performance 🤝 developer ergonomics.

Full results: https://huggingface.co/spaces/treble-technologies/ffasr

Want to learn more about the Far-Field ASR benchmark?

Join Treble's webinar tomorrow, June 11th, with @shinjiw_at_cmu, Cohere's @Julianfmack, and other industry leaders discussing the future of far-field speech recognition.

Register here: https://www.treble.tech/insights/treble-hugging-face-ffasr-webinar

@nickfrosst Nice, keeping that momentum after 30b.

@cohere @huggingface You folks are cooking! Thanks for OSS'ing it. We will add to models we host

@cohere @victormustar @huggingface Can it make word level timestamps or not? If not what is the trained durations that it will generate sentence longs

Block text not useful as subtitle

@nickfrosst Glad to see that Cohere is dropping small open sourced models like this

@cohere @huggingface Might try this out to help a call center in Quebec, hope it will work well in French. What kind of hardware would be required to have near real-time translation for, let's say, 10 to 20 users? And if we use the API, can you guarantee the confidential handling of information?

@cohere @huggingface And how much did this cost Canadian taxpayers to build?

the margin grows as conditions get worse. in low snr cohere is ~4 WER points ahead of ibm, but in near field theyre basically tied. the asymmetric encoder/decoder split makes sense for this: most compute goes to acoustic features, and the light decoder keeps latency down. curious how this runs on device vs cloud

@nickfrosst kind of satisfying when the metric catches up to the work already done.

proof you built the right thing, not the trendy thing.

@nickfrosst undetected ecov did transcribe audio prior and the apache 2.0 is more stronger