Cohere Labs drops 2B ASR model

// 75d agoMODEL RELEASE

Cohere Labs drops 2B ASR model

Cohere Labs released Cohere Transcribe, an open-source 2B-parameter speech-to-text model on Hugging Face that supports 14 languages. It uses a Conformer encoder and Transformer decoder, and it is aimed at offline transcription and production serving via transformers or vLLM.

// ANALYSIS

My read: Cohere is treating ASR like serious infrastructure, not a demo feature. The release is strong on accuracy and deployment ergonomics, but it still leaves common transcript workflows to external tooling.

–Cohere claims best-in-class transcription accuracy and up to 3x faster real-time factor than other dedicated ASR models in the same size range.
–The model supports 14 languages, but it does not auto-detect language and performs poorly on code-switched audio.
–It also ships without timestamps or speaker diarization, so meeting and call analytics stacks will still need extra layers.
–The repo is Apache 2.0, but Hugging Face adds a small friction point because users must accept contact-info sharing before downloading.

// TAGS

cohere-transcribespeechopen-sourceinferencebenchmark

DISCOVERED

75d ago

2026-03-26

PUBLISHED

75d ago

2026-03-26

RELEVANCE

9/ 10

AUTHOR

LinkSea8324

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

OPEN SOURCE5m ago

OpenMed 1.5.5 adds batch PII processing

OpenMed version 1.5.5 introduces batch PII extraction and de-identification alongside REST model unload and keep-alive controls. The update also improves Swift-based on-device acceleration and clinical OCR document handling on Apple Silicon.

OPEN SOURCE12m ago

A new open-source repository, train-llm-from-scratch, provides a step-by-step guide to building and training a transformer model from raw data on a single GPU.

The train-llm-from-scratch repository by Fareed Khan provides an end-to-end, open-source guide to building a transformer model from scratch. Unlike typical tutorials that focus solely on the neural network architecture, this project walks developers through the entire pipeline: downloading and parsing raw text data, implementing tokenization, structuring the transformer layers, training the model on a single GPU, and running inference. It aims to make the underlying mechanics of modern LLMs accessible and practical for individual developers.

NEWS15m ago

Claude Code creator highlights shift to agentic loops

Boris Cherny, the creator of Anthropic's Claude Code, shared that his engineering workflow has shifted from writing traditional prompts to building autonomous loops. Developers now build agentic systems that run continuously to discover, execute, and verify tasks independent of constant user input.

Cohere Labs drops 2B ASR model