Deep-unlearning

Deep-unlearning / nano-cohere-transcribe

Public

Pure-PyTorch inference for CohereLabs/cohere-transcribe-03-2026 (2B Conformer + Transformer ASR, 14 languages).

19
1
100% credibility
Found Apr 30, 2026 at 19 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

A lightweight tool for fast, accurate speech-to-text transcription in 14 languages using a Conformer-Transformer model.

How It Works

1
👀 Discover the transcriber

You hear about a speedy tool that turns audio recordings into text for 14 languages like English, French, or Japanese.

2
📦 Get it ready

You set up the lightweight transcriber on your computer, agreeing to the AI model's sharing rules.

3
🎤 Pick your audio

Choose an audio file from a short clip or a long meeting, and select the spoken language.

4
🚀 Transcribe automatically

Hit go and watch it smartly split long recordings at quiet spots, processing everything blazing fast.

5
⚙️ Tweak for speed

Adjust how many pieces it handles at once to balance speed and your computer's power.

✅ Read your transcript

Enjoy accurate, clean text output you can copy, edit, or use right away.

Sign up to see the full architecture

4 more

Sign Up Free

Star Growth

See how this repo grew from 19 to 19 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is nano-cohere-transcribe?

nano-cohere-transcribe is a lightweight Python package for pure-PyTorch inference on CohereLabs/cohere-transcribe-03-2026, a 2B-parameter Conformer-Transformer ASR model supporting 14 languages like English, French, Arabic, and Japanese. It lets you transcribe audio via a simple CLI (`nano-cohere-transcribe audio.wav --language en`) or Python API (`model.transcribe(waveform)`), handling short clips or long-form files with automatic energy-based chunking. Ditch the heavy transformers library and its import bloat for a minimal runtime with just six dependencies.

Why is it gaining traction?

It delivers 1.5x-3.6x faster inference than the native transformers path on benchmarks like earnings calls, matching leaderboard WER (10.86% on earnings22) while using less VRAM via batched chunk processing and CUDA graphs. Developers love the tiny footprint—no accelerate or generate scaffolding—and plug-and-play detokenization for HF pipelines. Greedy decoding keeps it simple and fast for production transcription.

Who should use this?

Backend devs building real-time ASR pipelines in Python apps, especially for multilingual podcasts, earnings calls, or voice apps needing 14 languages. ML engineers deploying Cohere's model without transformers overhead, or anyone processing long audio where speed and low deps matter over beam search.

Verdict

Grab it if you need fast, lightweight Cohere ASR—docs, benchmarks, and tests are solid for a 19-star project. At 1.0% credibility, it's early beta; production users should monitor for edge cases before scaling.

(187 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.