Cohere Transcribe CLI

Speaker-diarized transcription CLI powered by Cohere Transcribe and pyannote

Prerequisites

• Python 3.12+
• uv
• ffmpeg (required for YouTube downloads)

Setup

1. Install dependencies

uv sync

2. Authenticate with Hugging Face

Both the Cohere Transcribe and pyannote speaker-diarization models are gated. You must:

1. Accept the terms on each model page linked above.
2. Log in with the Hugging Face CLI:

uvx hf auth login

This will prompt you for a User Access Token with read access.

Usage

Transcribe a local audio file

uv run python main.py --audio recording.wav

Transcribe from YouTube

uv run python main.py --youtube "https://www.youtube.com/watch?v=VIDEO_ID"

Options

Flag	Default	Description
--audio		Path to a local audio file
--youtube		YouTube URL to download and transcribe
--language	en	Language code (e.g. en, fr, de, es, ja, zh)
--num-speakers	auto	Fixed number of speakers (auto-detected if omitted)
--backend	auto	ASR backend: mlx, cuda, cpu, or auto
--output	transcription.txt	Output file path
--merge-gap	0.35	Max gap (seconds) to merge same-speaker segments
--min-island	0.20	Min duration (seconds) for isolated speaker segments
--left-pad	0.35	Padding before each segment (seconds)
--right-pad	0.05	Padding after each segment (seconds)

Backend selection

By default (--backend auto), the CLI picks the best available backend:

• MLX on Apple silicon
• CUDA on systems with an NVIDIA GPU
• CPU otherwise

You can override with --backend mlx, --backend cuda, or --backend cpu. Passing --backend cpu also forces diarization to run on CPU.

Example

uv run python main.py --audio meeting.wav --num-speakers 3 --output meeting.txt

Output (meeting.txt):

SPEAKER_00 [00:00.00 - 00:12.34]:
Welcome everyone to the meeting. Let's start with the first agenda item.

SPEAKER_01 [00:12.34 - 00:25.67]:
Thanks. I wanted to discuss the timeline for the next release.