WER measured on the full LibriSpeech test-clean split (2620 utterances) with greedy decoding and no external LM. BF16 reference baseline: 1.26%. Cohere's self-reported number on the same split is 1.25% (Open ASR Leaderboard, as of 2026-03-26). Both ours and Cohere's numbers use the Whisper EnglishTextNormalizer, so the comparison is apples-to-apples and our port matches the upstream reference within rounding.

Usage

Build transcribe.cpp from source:

git clone git@github.com:handy-computer/transcribe.cpp.git
cd transcribe.cpp
cmake -B build && cmake --build build

Run on a 16 kHz mono WAV:

build/bin/transcribe-cli \
  -m cohere-transcribe-03-2026-Q8_0.gguf \
  input.wav

If your audio isn't already 16 kHz mono WAV, convert it first:

ffmpeg -i input.mp3 -ar 16000 -ac 1 output.wav

See the transcribe.cpp model page for performance numbers, numerical validation, and reproduction steps.

License

Inherited from the base model: Apache-2.0. See the upstream model card for full terms.

Original Model Card

The section below is reproduced from CohereLabs/cohere-transcribe-03-2026 at commit 76b8b23 for offline reference. The upstream card is the authoritative source.

Cohere Transcribe

Cohere Transcribe is an open source release of a 2B parameter dedicated audio-in, text-out automatic speech recognition (ASR) model. The model supports 14 languages.

Developed by: Cohere and Cohere Labs. Point of Contact: Cohere Labs.

✨Try the Cohere Transcribe demo✨

Usage

Cohere Transcribe is supported natively in transformers. This is the recommended way to use the model for offline inference. For online inference, see the vLLM integration example below.

pip install transformers>=5.4.0 torch huggingface_hub soundfile librosa sentencepiece protobuf
pip install datasets  # only needed for long-form and non-English examples

Testing was carried out with torch==2.10.0 but it is expected to work with other versions.

Quick Start 🤗

Transcribe any audio file in a few lines:

from transformers import AutoProcessor, CohereAsrForConditionalGeneration
from transformers.audio_utils import load_audio
from huggingface_hub import hf_hub_download

processor = AutoProcessor.from_pretrained("CohereLabs/cohere-transcribe-03-2026")
model = CohereAsrForConditionalGeneration.from_pretrained("CohereLabs/cohere-transcribe-03-2026", device_map="auto")

audio_file = hf_hub_download(
    repo_id="CohereLabs/cohere-transcribe-03-2026",
    filename="demo/voxpopuli_test_en_demo.wav",
)
audio = load_audio(audio_file, sampling_rate=16000)

inputs = processor(audio, sampling_rate=16000, return_tensors="pt", language="en")
inputs.to(model.device, dtype=model.dtype)

outputs = model.generate(**inputs, max_new_tokens=256)
text = processor.decode(outputs, skip_special_tokens=True)
print(text)

For audio longer than the feature extractor's max_audio_clip_s, the feature extractor automatically splits the waveform into chunks. The processor reassembles the per-chunk transcriptions using the returned audio_chunk_index.

This example transcribes a 55 minute earnings call:

from transformers import AutoProcessor, CohereAsrForConditionalGeneration
from datasets import load_dataset
import time

processor = AutoProcessor.from_pretrained("CohereLabs/cohere-transcribe-03-2026")
model = CohereAsrForConditionalGeneration.from_pretrained("CohereLabs/cohere-transcribe-03-2026", device_map="auto")

ds = load_dataset("distil-whisper/earnings22", "full", split="test", streaming=True)
sample = next(iter(ds))

audio_array = sample["audio"]["array"]
sr = sample["audio"]["sampling_rate"]
duration_s = len(audio_array) / sr
print(f"Audio duration: {duration_s / 60:.1f} minutes")

inputs = processor(audio=audio_array, sampling_rate=sr, return_tensors="pt", language="en")
audio_chunk_index = inputs.get("audio_chunk_index")
inputs.to(model.device, dtype=model.dtype)

start = time.time()
outputs = model.generate(**inputs, max_new_tokens=256)
text = processor.decode(outputs, skip_special_tokens=True, audio_chunk_index=audio_chunk_index, language="en")[0]
elapsed = time.time() - start
rtfx = duration_s / elapsed
print(f"Transcribed in {elapsed:.1f}s — RTFx: {rtfx:.1f}")
print(f"Transcription ({len(text.split())} words):")
print(text[:500] + "...")

Pass punctuation=False to obtain lower-cased output without punctuation marks.

inputs_pnc = processor(audio, sampling_rate=16000, return_tensors="pt", language="en", punctuation=True)
inputs_nopnc = processor(audio, sampling_rate=16000, return_tensors="pt", language="en", punctuation=False)

By default, punctuation is enabled.

Multiple audio files can be processed in a single call. When the batch mixes short-form and long-form audio, the processor handles chunking and reassembly.

from transformers import AutoProcessor, CohereAsrForConditionalGeneration
from transformers.audio_utils import load_audio

processor = AutoProcessor.from_pretrained("CohereLabs/cohere-transcribe-03-2026")
model = CohereAsrForConditionalGeneration.from_pretrained("CohereLabs/cohere-transcribe-03-2026", device_map="auto")

audio_short = load_audio(
    "https://huggingface.co/datasets/hf-internal-testing/dummy-audio-samples/resolve/main/bcn_weather.mp3",
    sampling_rate=16000,
)
audio_long = load_audio(
    "https://huggingface.co/datasets/hf-internal-testing/dummy-audio-samples/resolve/main/obama_first_45_secs.mp3",
    sampling_rate=16000,
)

inputs = processor([audio_short, audio_long], sampling_rate=16000, return_tensors="pt", language="en")
audio_chunk_index = inputs.get("audio_chunk_index")
inputs.to(model.device, dtype=model.dtype)

outputs = model.generate(**inputs, max_new_tokens=256)
text = processor.decode(
    outputs, skip_special_tokens=True, audio_chunk_index=audio_chunk_index, language="en"
)
print(text)

Specify the language code to transcribe in any of the 14 supported languages. This example transcribes Japanese audio from the FLEURS dataset:

from transformers import AutoProcessor, CohereAsrForConditionalGeneration
from datasets import load_dataset

processor = AutoProcessor.from_pretrained("CohereLabs/cohere-transcribe-03-2026")
model = CohereAsrForConditionalGeneration.from_pretrained("CohereLabs/cohere-transcribe-03-2026", device_map="auto")

ds = load_dataset("google/fleurs", "ja_jp", split="test", streaming=True)
ds_iter = iter(ds)
samples = [next(ds_iter) for _ in range(3)]

for sample in samples:
    audio = sample["audio"]["array"]
    sr = sample["audio"]["sampling_rate"]

    inputs = processor(audio, sampling_rate=sr, return_tensors="pt", language="ja")
    inputs.to(model.device, dtype=model.dtype)

    outputs = model.generate(**inputs, max_new_tokens=256)
    text = processor.decode(outputs, skip_special_tokens=True)
    print(f"REF: {sample['transcription']}\nHYP: {text}\n")

Broader dependency support with trust_remote_code=True

For a wider range of torch and transformers versions, run with trust_remote_code=True. You should expect greater stability via the transformers native path above. This option will be deprecated in the future.

trust_remote_code=True inference exposes a single model.transcribe() method that automatically handles long-form audio chunking and exposes parameters to facilitate efficient inference. It is recommended that you let the transcribe method handle batching for you. This implementation is optimized for offline inference: for online inference, see the vLLM integration example below.

Installation

Recommended:

pip install "transformers>=4.56,<5.3,!=5.0.*,!=5.1.*" torch huggingface_hub soundfile librosa sentencepiece protobuf
pip install datasets  # only needed for examples 2 and 3

For broader compatibility run the following install:

pip install "transformers>=4.52,!=5.0.*,!=5.1.*" torch huggingface_hub soundfile librosa sentencepiece protobuf

This will replace the efficient static-cache with a dynamic-cache fallback on some versions.

Transformers 5.0 and 5.1 have a weight-loading issue and are not compatible.

Example 1: Quick Start

Transcribe any audio file in a few lines. The model accepts file paths directly — no manual preprocessing required.

import torch
from transformers import AutoProcessor, AutoModelForSpeechSeq2Seq
from huggingface_hub import hf_hub_download

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)
model.eval()

audio_file = hf_hub_download(
    repo_id="CohereLabs/cohere-transcribe-03-2026",
    filename="demo/voxpopuli_test_en_demo.wav",
)

texts = model.transcribe(processor=processor, audio_files=[audio_file], language="en")
print(texts[0])

When audio is already in memory (streaming datasets, microphone input, etc.), pass numpy arrays directly instead of file paths. Enable compile=True to torch.compile the encoder for faster throughput, and pipeline_detokenization=True to overlap CPU detokenization with GPU inference.

Note: pipeline_detokenization=True is not supported on Windows.

This example transcribes Japanese audio from the FLEURS dataset:

import torch
from transformers import AutoProcessor, AutoModelForSpeechSeq2Seq
from huggingface_hub import hf_hub_download


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)
model.eval()

from datasets import load_dataset

ds = load_dataset("google/fleurs", "ja_jp", split="test", streaming=True)
ds_iter = iter(ds)
samples = [next(ds_iter) for _ in range(3)]  # take 3 samples

audio_arrays = [s["audio"]["array"] for s in samples]
sample_rates = [s["audio"]["sampling_rate"] for s in samples]

# compile=True incurs a one-time warmup cost on the first call; subsequent calls are faster.
texts = model.transcribe(
    processor=processor,
    audio_arrays=audio_arrays,
    sample_rates=sample_rates,
    language="ja",
    compile=True,
    pipeline_detokenization=True,
    batch_size=16,
)
for ref, hyp in zip([s["transcription"] for s in samples], texts):
    print(f"REF: {ref}\nHYP: {hyp}\n")

Audio longer than 35 seconds is automatically split into overlapping chunks and reassembled. The API is identical — no special flags or configuration needed. This example transcribes a 55 minute earnings call. This will be slow if you haven't run compile=True in the previous example:

import torch
from transformers import AutoProcessor, AutoModelForSpeechSeq2Seq
from huggingface_hub import hf_hub_download


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)
model.eval()

from datasets import load_dataset

ds = load_dataset("distil-whisper/earnings22", "full", split="test", streaming=True)
sample = next(iter(ds))

import time

audio_array = sample["audio"]["array"]
sr = sample["audio"]["sampling_rate"]
duration_s = len(audio_array) / sr
print(f"Audio duration: {duration_s / 60:.1f} minutes")

start = time.time()
texts = model.transcribe(
    processor=processor,
    audio_arrays=[audio_array],
    sample_rates=[sr],
    language="en",
    compile=True,
)
elapsed = time.time() - start
rtfx = duration_s / elapsed
print(f"Transcribed in {elapsed:.1f}s — RTFx: {rtfx:.1f}")
print(f"Transcription ({len(texts[0].split())} words):")
print(texts[0][:500] + "...")

Returns: list[str] — one transcription string per input audio.

vLLM Integration

For production serving we recommend running via vLLM following the instructions below.

First install vLLM (refer to vLLM installation instructions):

uv pip install -U vllm --torch-backend=auto --extra-index-url https://wheels.vllm.ai/nightly
uv pip install vllm[audio]
uv pip install librosa

Start vLLM server

vllm serve CohereLabs/cohere-transcribe-03-2026 --trust-remote-code

Send request

curl -v -X POST http://localhost:8000/v1/audio/transcriptions \
 -H "Authorization: Bearer $VLLM_API_KEY" \
-F "file=@$(realpath ${AUDIO_PATH})" \
-F "model=CohereLabs/cohere-transcribe-03-2026"

Results

Link to the live leaderboard: Open ASR Leaderboard.

Human-preference results

We observe similarly strong performance in human evaluations, where trained annotators assess transcription quality across real-world audio for accuracy, coherence and usability. The consistency between automated metrics and human judgments suggests that the model’s improvements translate beyond controlled benchmarks to practical transcription settings.

Figure: Human preference evaluation of model transcripts. In a head-to-head comparison, annotators were asked to express preferences for generations which primarily preserved meaning - but also avoided hallucination, correctly identified named entities, and provided verbatim transcripts with appropriate formatting. A score of 50% or higher indicates that Cohere Transcribe was preferred on average in the comparison.

Figure: per-language error rate averaged over FLEURS, Common Voice 17.0, MLS and Wenet tests sets (where relevant for a given language). CER for zh, ja, ko — WER otherwise

Resources

For more details and results:

• Technical blog post contains WERs and other quality metrics.
• Announcement blog post for more information about the model.
• English, EU and long-form transcription WERs/RTFx are on the Open ASR Leaderboard.

Strengths and Limitations

Cohere Transcribe is a performant, dedicated ASR model intended for efficient speech transcription.

Strengths

Cohere Transcribe demonstrates best-in-class transcription accuracy in 14 languages. As a dedicated speech recognition model, it is also efficient, benefitting from a real-time factor up to three times faster than that of other, dedicated ASR models in the same size range. The model was trained from scratch, and from the outset, we deliberately focused on maximizing transcription accuracy while keeping production readiness top-of-mind.

Limitations

• Single language. The model performs best when remaining in-distribution of a single, pre-specified language amongst the 14 in the range it supports. It does not feature explicit, automatic language detection and exhibits inconsistent performance on code-switched audio.

• Timestamps/Speaker diarization. The model does not feature either of these.

• Silence. Like most AED speech models, Cohere Transcribe is eager to transcribe, even non-speech sounds. The model thus benefits from prepending a noise gate or VAD (voice activity detection) model in order to prevent low-volume, floor noise from turning into hallucinations.

Ecosystem support 🚀

Cohere Transcribe is supported on the following libraries/platforms:

• transformers (see Quick Start above).
• vLLM (see vLLM integration above).
• mlx-audio for Apple Silicon.
• Rust implementation: cohere_transcribe_rs
• In the browser ✨demo✨ (via transformers.js and WebGPU)
• Chrome extension: cohere_transcribe_extension
• Whisper Memos (iOS App).
• Whisperian (Android App).

If you have added support for the model somewhere not included above please raise an issue/PR!

If you find issues with any of these please raise an issue with the respective library.

Model Card Contact

For errors or additional questions about details in this model card, contact labs@cohere.com or raise an issue.

Terms of Use: We hope that the release of this model will make community-based research efforts more accessible, by releasing the weights of a highly performant 2 billion parameter model to researchers all over the world. This model is governed by an Apache 2.0 license.