Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/KoljaB/RealtimeSTT/llms.txt

Use this file to discover all available pages before exploring further.

The Cohere Transcribe engine loads CohereLabs/cohere-transcribe-03-2026 locally through Hugging Face Transformers. This is an experimental adapter — it does not call a hosted API and does not use the openai_api placeholder. Because the model does not auto-detect language, an explicit language code is always required.
Cohere Transcribe does not auto-detect language. You must pass language="en" (or another supported code) or the engine will raise an error. This applies whether you set the language via the top-level language parameter or through transcription_engine_options["language"].

​
Install

The cohere extra is an alias for the Transformers backend. Install it with: 
pip install "RealtimeSTT[cohere]"
Or install the dependencies manually: 
pip install transformers torch
Install any model-specific dependencies required by your installed Transformers release. The model may also require accepting gated Hugging Face access before weights can be downloaded.

​
Engine Names

Both of the following names are accepted as the transcription_engine value:
  • cohere_transcribe
  • cohere

​
Basic Usage

from RealtimeSTT import AudioToTextRecorder

recorder = AudioToTextRecorder(
    transcription_engine="cohere_transcribe",
    model="CohereLabs/cohere-transcribe-03-2026",
    language="en",
    device="cuda",
)
You can also supply the language through transcription_engine_options instead of the top-level language parameter: 
recorder = AudioToTextRecorder(
    transcription_engine="cohere",
    model="CohereLabs/cohere-transcribe-03-2026",
    transcription_engine_options={"language": "en"},
)

​
Hugging Face Model Access

This adapter loads the model locally through Transformers. If the model is gated, you must authenticate with a Hugging Face token before weights can be downloaded. Use one of the following methods: 
huggingface-cli login
Or set the environment variable before running your script: 
export HF_TOKEN=hf_your_token_here

​
Custom Model Cache Directory

Set download_root to map model and processor cache files to a project-local directory: 
recorder = AudioToTextRecorder(
    transcription_engine="cohere_transcribe",
    model="CohereLabs/cohere-transcribe-03-2026",
    language="en",
    download_root="models/hf",
)

​
Configuration Options

OptionMeaning
engine_options["processor"]Passed to AutoProcessor.from_pretrained.
engine_options["model"]Passed to CohereAsrForConditionalGeneration.from_pretrained.
engine_options["processor_call"]Merged into the processor call.
engine_options["generate"]Merged into model.generate(...).
engine_options["decode"]Merged into processor decode.
engine_options["language"]Language when not passed through the top-level language parameter.
engine_options["punctuation"]Passed to the processor when set.
engine_options["sample_rate"]Input sample rate. Defaults to 16000.

​
Latency and Cost Notes

Because this adapter runs a local model, runtime cost is local compute rather than per-request API billing. The practical cost is model download size, startup time, GPU or CPU memory, and inference latency. Treat it as a heavier experimental backend until you have validated it on your target hardware.

​
Troubleshooting

The engine raises an error if neither language nor engine_options["language"] is set. Pass a language code such as language="en". Cohere Transcribe does not auto-detect language.
Accept the model’s access terms on the Hugging Face model page, then authenticate using huggingface-cli login or the HF_TOKEN environment variable.
A AttributeError for CohereAsrForConditionalGeneration or AutoProcessor usually means the installed Transformers version is too old. Upgrade to a recent release: pip install --upgrade transformers.

Build docs developers (and LLMs) love

Get started for freeTalk to us