Cohere Transcribe App

A Streamlit application that transcribes audio locally using the Cohere Transcribe ONNX model (2B parameters, #1 on HF Open ASR Leaderboard).

Architecture

Audio File / URL
       |
       v
  librosa (load + resample to 16kHz)
       |
       v
  CohereAsrProcessor (mel spectrogram + prompt tokens)
       |
       v
  ONNX Encoder (Conformer) --> encoder hidden states
       |
       v
  ONNX Decoder (autoregressive greedy + KV cache) --> token IDs
       |
       v
  Tokenizer decode --> transcript text

Long audio is split into 30s chunks with 5s overlap, each chunk processed independently.

Setup

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
pip install git+https://github.com/huggingface/transformers.git
streamlit run app.py

Note: transformers must be installed from source for cohere_asr model support.

Supported Languages

Arabic, Chinese, Dutch, English, French, German, Greek, Italian, Japanese, Korean, Polish, Portuguese, Spanish, Vietnamese

Model

• Source: CohereLabs/cohere-transcribe-03-2026
• ONNX: onnx-community/cohere-transcribe-03-2026-ONNX
• Quantization options: INT8, Q4, FP16
• Runs on CPU (~3.5 GB RAM)