- Updated: March 26, 2026
- 3 min read
Cohere Launches Open‑Source Transcribe Model for Multilingual Speech Recognition
**Cohere Launches “Transcribe” – an Open‑Source ASR Model**
| Aspect | Details |
|——–|———|
| **Product** | **Transcribe** – Cohere’s first voice‑model, an automatic‑speech‑recognition (ASR) system released as open‑source software. |
| **Size & Compute** | 2 billion parameters – deliberately lightweight so it can run on consumer‑grade GPUs (e.g., RTX 3060/3070) and be self‑hosted. |
| **Language Coverage** | 14 languages: English, French, German, Italian, Spanish, Portuguese, Greek, Dutch, Polish, Chinese (Mandarin), Japanese, Korean, Vietnamese, Arabic. |
| **Performance** | *Hugging Face Open ASR leaderboard*: WER = 5.42 % – the lowest among all listed models.
*Human evaluation*: 61 % win‑rate vs. competing systems on accuracy, coherence, and usability.
*Weak spots*: lower relative performance on Portuguese, German, and Spanish. |
| **Speed** | Claims to transcribe **525 minutes of audio per minute** (≈ 8.75 × real‑time), which is high for a 2 B‑parameter model. |
| **Benchmarks & Competitors** | Beats Zoom Scribe v1, IBM Granite 4.0 1B, ElevenLabs Scribe v2, and Qwen3‑ASR‑1.7B on the Open ASR leaderboard. |
| **Availability** | • **GitHub / Model Hub** – full model weights and inference code are open‑source.
• **Cohere API** – free tier access (no charge for usage).
• **Model Vault** – Cohere’s managed inference platform for those who prefer a hosted solution. |
| **Planned Integration** | Will be embedded in Cohere’s enterprise agent‑orchestration platform **North**, enabling downstream applications (e.g., note‑taking bots, call‑center analytics) to call Transcribe directly. |
| **Business Context** | • Cohere reported **$240 M ARR** for 2025 and hinted at a possible IPO “soon”.
• The launch aligns with a broader market surge in speech‑AI (note‑taking apps like Granola, Wispr Flow).
• By open‑sourcing a competitive model, Cohere positions itself as a community‑first AI infrastructure provider while still monetizing via managed services (Model Vault, North). |
| **Nuances & Caveats** | • **Open‑source vs. proprietary**: While the model is free, Cohere may capture value through premium support, managed hosting, and integration with its paid North platform.
• **Language gaps**: The lower performance on Portuguese, German, and Spanish suggests further fine‑tuning or larger multilingual models may be needed for enterprise customers with heavy usage in those languages.
• **Speed claim**: 525 min/min is measured on high‑end GPUs; real‑world throughput on typical consumer hardware will be lower, though still well above real‑time.
• **Evaluation methodology**: Human win‑rate (61 %) is based on a limited evaluator pool; the metric complements but does not replace standard WER. |
| **Related News & Events** | • The article appears in a TechCrunch piece by Ivan Mehta (published March 26 2026).
• Mention of upcoming TechCrunch Disrupt 2026 and Founder Summit events (unrelated to the model but part of the broader tech‑news context). |
### TL;DR Summary
Cohere has released **Transcribe**, a **2 B‑parameter, open‑source ASR model** that runs on consumer GPUs and supports **14 languages**. It achieves a **5.42 % word‑error rate**, the best score on the Hugging Face Open ASR leaderboard, and outperforms several commercial rivals in human‑rated accuracy. The model can process **525 minutes of audio per minute**, is freely available via GitHub, Cohere’s API, and its managed Model Vault, and will be integrated into Cohere’s enterprise orchestration platform **North**. While strong overall, it lags slightly on Portuguese, German, and Spanish. The launch dovetails with Cohere’s rapid growth (≈ $240 M ARR in 2025) and its strategy of offering free, community‑driven models while monetizing through hosted services and enterprise tooling.
Read the full story on TechCrunch. Explore related insights on our site: Cohere Transcribe Overview and AI Model Updates.
Andrii Bidochko
CTO UBOS
Andrii Bidochko is an AI entrepreneur and researcher focused on AI agents, reinforcement learning, and autonomous systems. He writes about the technologies shaping the future of machine intelligence, from frontier models and agent architectures to real-world AI applications.