---
license: apache-2.0
language:
- ar
- de
- el
- en
- es
- fr
- it
- ja
- ko
- nl
- pl
- pt
- vi
- zh
pipeline_tag: automatic-speech-recognition
tags:
- audio
- hf-asr-leaderboard
- speech-recognition
- transcription
- openvino
library_name: transformers
---

# Cohere Transcribe OpenVINO

This repository contains an **OpenVINO™ IR** version of Cohere Transcribe optimized for efficient inference with the OpenVINO Runtime. The model has been exported into separate **encoder**, **decoder**, and **decoder-with-past** graphs to enable fast autoregressive decoding using KV-cache.

The repository includes everything required to run inference with OpenVINO, including the model IR files, tokenizer, processor configuration, and metadata.

## Model Details

* **Task:** Automatic Speech Recognition (ASR)
* **Framework:** OpenVINO Runtime
* **Model Format:** OpenVINO IR (`.xml` + `.bin`)
* **Supported Devices:** CPU, GPU, AUTO
* **Language:** English, 

---

# Repository Contents

```
.
├── encoder.xml
├── encoder.bin
├── decoder.xml
├── decoder.bin
├── decoder_with_past.xml
├── decoder_with_past.bin
├── config.json
├── generation_config.json
├── preprocessor_config.json
├── tokenizer.json
├── tokenizer_config.json
├── special_tokens_map.json
├── meta.json
└── README.md
```

---

# Installation

We recommend using **uv** to create an isolated Python environment.

## 1. Install uv

```bash
pip install uv
```

## 2. Create a virtual environment

```bash
uv venv
```

Activate the environment.

### Linux / macOS

```bash
source .venv/bin/activate
```

### Windows (PowerShell)

```powershell
.venv\Scripts\Activate.ps1
```

### Windows (CMD)

```cmd
.venv\Scripts\activate.bat
```

---

## Install Dependencies

```bash
uv pip install \
    openvino \
    transformers \
    numpy \
    librosa \
    datasets
```

or using pip

```bash
pip install \
    openvino \
    transformers \
    numpy \
    librosa \
    datasets
```

---

# Quick Start

```python
python inference.py
```

---

# Supported Devices

The model can be executed on any OpenVINO-supported device.

| Device       | Supported |
| ------------ | --------- |
| CPU          | ✅         |
| GPU          | ✅         |
| AUTO         | ✅         |
| AUTO:CPU,GPU | ✅         |

---

# Performance Notes

The model uses:

* OpenVINO Runtime
* Separate encoder and decoder graphs
* KV-cache based decoding
* Decoder-with-past for efficient autoregressive generation

These optimizations significantly reduce decoding latency compared to running the full decoder at every generation step.

---

# Acknowledgements

This model is an OpenVINO conversion of the original Cohere transcribe model.

Special thanks to:

* **OpenVINO™ Toolkit** for providing the runtime and model optimization tools.
* **Hugging Face Transformers** for preprocessing, tokenization, and model configuration support.
* **Cohere Labs** for the original cohere-transcribe-03-2026 model.

Useful resources:

* OpenVINO Documentation: https://docs.openvino.ai
* OpenVINO GitHub: https://github.com/openvinotoolkit/openvino
* Hugging Face OpenVINO Documentation: https://huggingface.co/docs/optimum/intel/openvino

---

# License

This repository contains **only an OpenVINO IR conversion** of the original model.

Please refer to the **original Cohere transcribe model repository** for the model license and usage terms. The conversion itself does not modify the licensing of the original model.

---