Cohere Transcribe Tops Every Noisy Real-World Speech Benchmark at 525x Speed
Cohere Transcribe claims #1 on the new Far-Field ASR benchmark, beating IBM and NVIDIA with a 17.9 WER on noisy, real-world audio
- Double leaderboard win: Cohere Transcribe is now #1 on both the Hugging Face Open ASR Leaderboard (5.42 WER) and the new Far-Field ASR benchmark (17.9 WER).
- New FFASR benchmark: Treble Technologies and Hugging Face launched the first open benchmark testing ASR models under real-world noisy, reverberant, far-field conditions.
- Beats IBM and NVIDIA: Transcribe finished ~2 WER points ahead of IBM Granite Speech and ~3.6 points ahead of NVIDIA Parakeet on FFASR.
- Architecture edge: A 2B Fast-Conformer encoder with a lightweight decoder gives it 525x real-time speed — 3x faster than comparable models — while keeping accuracy best-in-class.
- Free and open: Apache 2.0 license, runs on consumer GPUs, available via Hugging Face with native Transformers and vLLM support.
- Known gaps: No automatic language detection, no speaker diarization, and weaker performance on Portuguese, German, and Spanish.
Cohere's open-source speech recognition model just pulled off a double: after topping the Hugging Face Open ASR Leaderboard for clean audio back in March, Cohere Transcribe has now claimed the #1 spot on the brand-new Far-Field ASR (FFASR) Leaderboard , a benchmark specifically designed to stress-test models in the messy acoustic conditions where enterprise voice AI actually lives.
A benchmark built for the real world
Most ASR benchmarks measure how well a model handles clean, close-microphone audio. That is not how meetings, contact centers, or phone calls sound. Standard ASR benchmarks often miss the conditions that matter most in real-world use: background noise, competing speech, reverberation, distance from the microphone, and other acoustic effects.
Treble Technologies and Hugging Face launched the Far-Field ASR (FFASR) Leaderboard , the first open, community-driven benchmark designed to evaluate ASR models under realistic far-field acoustic conditions. Far-field audio is what you get when the speaker is not right next to the microphone: think a person talking across a conference table, or a customer calling in over a noisy phone line. The benchmark tests models across varying signal-to-noise ratios (SNR , the ratio of the desired speech signal to background noise) and reverberant environments (rooms where sound bounces off walls and creates echo).
The numbers that matter
On the FFASR benchmark, Cohere Transcribe achieved a 17.9 WER (Word Error Rate , the percentage of words transcribed incorrectly, lower is better), finishing nearly 2 points ahead of IBM Granite Speech and 3.6 points ahead of NVIDIA's Parakeet. Critically, it ranked #1 across every metric on the leaderboard, not just the average.
This follows its earlier clean-audio win. Transcribe leads the Hugging Face Open ASR Leaderboard with an average word error rate of just 5.42%, outperforming all open- and closed-source dedicated ASR alternatives, including Whisper Large v3, ElevenLabs Scribe v2, and Qwen3-ASR-1.7B. The full leaderboard snapshot at launch looked like this:
| Model | Avg WER | AMI | LibriSpeech Clean | Earnings22 | RTFx |
|---|---|---|---|---|---|
| Cohere Transcribe | 5.42 | 8.15 | 1.25 | 10.84 | 525 |
| IBM Granite 4.0 1B Speech | 5.52 | 8.44 | 1.42 | 8.48 | 280 |
| NVIDIA Canary Qwen 2.5B | 5.63 | 10.19 | 1.61 | 10.45 | 418 |
| Qwen3-ASR-1.7B | 5.76 | 10.56 | 1.63 | 10.25 | 148 |
| OpenAI Whisper Large v3 | 7.44 | 15.95 | 2.01 | 11.29 | , |
Speed is part of the story too. With 2B parameters, it ranks #1 on the English ASR leaderboard, achieving a 5.42 average WER across 8 benchmarks, while running at 524x real-time speed (RTFx) , roughly 3x faster than comparable models. That means 525 minutes of audio transcribed per minute of compute, on a single GPU.
How it was built
The architecture is a deliberate bet on the encoder-decoder asymmetry. Cohere Transcribe is a 2B encoder-decoder transformer with a Fast-Conformer encoder. More than 90% of total parameters are dedicated to the encoder, with a lightweight decoder. This asymmetry keeps the amount of autoregressive inference compute to a minimum while maintaining performance , and is the direct reason for its efficiency advantage.
A Conformer (short for Convolutional Transformer) combines convolutional layers , good at capturing local acoustic patterns like phonemes , with self-attention layers that handle longer-range context like sentence rhythm. In contrast, models like Qwen-1.7B-ASR and IBM Granite 4.0 1B Speech build upon pre-trained text LLMs and add audio understanding to an autoregressive backbone. This makes the ASR model cheaper to train, but comes at the expense of inference speed and serving cost.
On the data side, Cohere took a high-volume, high-curation approach. The model was trained on 0.5 million hours of curated audio-transcript pairs, augmented with synthetic data following rounds of error analysis. During training, non-speech background noise augmentation was applied with signal-to-noise ratios in the range of 0 to 30 dB , which likely explains why the model holds up so well on the far-field benchmark's noisy conditions.
The real breakthrough: robustness, not just accuracy
The FFASR result is arguably more meaningful than the clean-audio leaderboard win. Any well-trained model can transcribe a podcast recorded in a studio. The harder problem , and the one enterprises actually face , is transcribing a sales call, a board meeting, or a voice assistant command from across the room.
For ASR models used in smart devices, meeting rooms, automotive systems, robotics, and other hands-free applications, clean near-field benchmarks only tell part of the story. FFASR evaluates models against realistic far-field acoustic scenarios, helping the community understand where models perform well, where they fail, and how they compare under conditions that are difficult to replicate in the lab.
The fact that Transcribe was trained with noise augmentation across a wide SNR range and still wins on a benchmark it was not explicitly tuned for suggests the gains are structural, not just benchmark-specific. Crucially, these gains are not limited to benchmark datasets. In Cohere's own human evaluations assessing accuracy, coherence, and usability, Transcribe achieved a 61% average win rate against competing models.
What it is bad at
No model is perfect, and Cohere is upfront about the gaps:
- No language detection. The model performs best when remaining in-distribution of a single, pre-specified language. It does not feature explicit, automatic language detection.
- No timestamps or speaker diarization. If you need to know who said what and when, you will need to layer another model on top.
- Hallucinations on silence. Like most encoder-decoder speech models, Transcribe will try to transcribe even non-speech sounds. Cohere recommends prepending a VAD (voice activity detection) model to filter out floor noise before it reaches the transcription model.
- Weaker on some languages. The model fell behind its rivals when transcribing Portuguese, German, and Spanish.
How to use it
The model is available now on Hugging Face under an Apache 2.0 license , no usage restrictions, commercially deployable. Getting started takes about five lines of Python:
from transformers import AutoProcessor, AutoModelForSpeechSeq2Seq
import torch
model_id = "CohereLabs/cohere-transcribe-03-2026"
device = "cuda:0" if torch.cuda.is_available() else "cpu"
processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForSpeechSeq2Seq.from_pretrained(model_id, trust_remote_code=True).to(device)
texts = model.transcribe(processor=processor, audio_files=["your_audio.wav"], language="en")
print(texts[0])For production serving, Cohere Transcribe is supported natively in the Hugging Face Transformers library for offline inference, with a vLLM integration available for online inference. The team contributed encoder-decoder batching improvements directly to vLLM, yielding up to 2x throughput gains for variable-length audio workloads. Automatic chunking for long-form audio, punctuation control, and broad ecosystem support across vLLM, Apple Silicon, browser, and mobile make it practical across a wide range of deployment scenarios.
The competitive landscape is moving fast
IBM released its Granite 4.0 1B Speech model just ten days before Transcribe's arrival, underscoring how rapidly the open-source ASR field is advancing. NVIDIA's Parakeet model had previously held the top ASR leaderboard position before Transcribe. The field is now so competitive that five speech recognition models sit within half a percentage point of each other on the Hugging Face Open ASR leaderboard.
Raw word error rate alone may no longer serve as a decisive differentiator for enterprises evaluating speech recognition. Selection criteria are increasingly shifting toward deployment flexibility, pricing, and multilingual coverage. The FFASR benchmark is a direct response to this: it forces models to prove themselves on the acoustic conditions that actually break production systems, not just the clean studio recordings that have historically defined the leaderboard.
For teams building meeting transcription, contact center analytics, or voice agents, developers gain a free alternative to commercial transcription services that now demonstrably holds up in the noisy, reverberant conditions those products ship into. The official release blog has the full architecture and training details.