Open-Source Speech Recognition

3 posts · Mar 2026 – Jul 2026

What this trail covers

Cohere has released Transcribe Arabic, an open-source, high-accuracy Automatic Speech Recognition (ASR) model. It is trained from scratch with a focus on minimizing Word Error Rate (WER) and is designed for practical enterprise use, offering manageable inference footprint and efficient serving. The model ranks #1 on HuggingFace's Open ASR Leaderboard and supports 14 languages. It can be downloaded from Hugging Face, accessed via API, or provisioned through Cohere's Model Vault for production deployment.

All Cohere trails

Jul 2026Tiny Aya Expedition Drives Multilingual Innovation | CohereThis post details the outcomes of the Expedition Tiny Aya program, highlighting projects that explored multilingual AI innovation across education, safety, accessibility, and language understanding. Key technical contributions include the development of a multilingual math solver with transparent reasoning, a child-safe offline AI voice assistant, investigations into multilingual safety under code-mixed prompts and cross-lingual misalignment, and research into cross-lingual semantic transfer. It also showcases advancements in local deployment through compression techniques, synthetic data pipelines, and privacy-preserving tools.
Jul 2026Cohere Transcribe Arabic: Open-Source Speech AI | CohereThis post details the release of Cohere Transcribe Arabic, an open-source ASR model specifically trained for the nuances of Arabic dialects, code-switching, and domain-specific vocabulary. It highlights its superior accuracy compared to existing models like Whisper and OmniASR, supported by detailed WER benchmarks across various datasets. The post also emphasizes its enterprise readiness through optimizations for high-throughput serving using vLLM, achieving a significant RTFx score. It provides examples demonstrating its effectiveness in handling dialectal variations and code-switching in workplace scenarios.
Mar 2026Cohere Transcribe: Open-Source Speech Recognition | CohereThis post details the release of Cohere Transcribe, an open-source Conformer-based encoder-decoder ASR model trained from scratch. It highlights its state-of-the-art accuracy (5.42% average WER) on the HuggingFace Open ASR Leaderboard, outperforming existing models. The post also discusses its performance across various datasets (AMI, Earnings 22, etc.) and human evaluations, emphasizing its robustness to real-world speech conditions. Furthermore, it presents throughput metrics (RTFx) demonstrating best-in-class serving efficiency for its model size cohort. The model supports 14 languages and is available via Hugging Face, API, and Model Vault.
Advertisement