TechPrufer
Sign inGet started

Blogs›Cohere›Open-Source Speech Recognition

Open-Source Speech Recognition

Open-Source Speech Recognition

3
posts
2026

Cohere has released Tiny Aya, an open-weight, lightweight multilingual model supporting over 70 languages, designed for local deployment and practical AI solutions. The 'Expedition Tiny Aya' program fostered innovation across education, safety, accessibility, and language understanding. Projects included a multilingual math solver with transparent reasoning, an offline child-safe AI voice assistant, investigations into multilingual safety under code-mixed prompts and cross-lingual misalignment, and the development of compression techniques, synthetic data pipelines, and privacy-preserving tools for local deployment. Research also demonstrated deep semantic cross-lingual transfer and sensitivity to programming language syntax.

2026

Tiny Aya Expedition Drives Multilingual Innovation | Cohere

7/14/2026

This post details the outcomes of the 'Expedition Tiny Aya' program, highlighting projects that leveraged the Tiny Aya model for multilingual innovation. Key technical contributions include: a multilingual math benchmark dataset, a child-safe AI voice assistant with a multilingual benchmark for child-focused AI, investigations into multilingual safety behavior under code-mixed prompts and cross-lingual misalignment, research into cultural context robustness in multilingual AI, a synthetic data generation pipeline for speech-to-speech translation in under-resourced languages, a multilingual retrieval and question-answering system for document understanding with on-device privacy, a parameter-efficient approach to connect visual encoders to Tiny Aya for accessible multilingual vision-language systems, and a system for on-device multilingual tool use via JSON tool calls with 4-bit quantized model variants.

Cohere Transcribe Arabic: Open-Source Speech AI | Cohere

7/7/2026

This post details the development and release of Cohere Transcribe Arabic, an extension of the existing Transcribe ASR model. It highlights the specific engineering challenges and solutions for handling Arabic's dialectical richness, code-switching, and bilingual speech. The post quantifies accuracy improvements over existing models (Whisper, OmniASR) using WER on various datasets and presents human evaluation results. It also details performance optimizations for production serving, including integration with vLLM for high-throughput workloads and RTFx scores.

Cohere Transcribe: Open-Source Speech Recognition | Cohere

3/26/2026

This post introduces Cohere Transcribe, a new open-source ASR model. It details the model's architecture (conformer-based encoder-decoder), training objective (supervised cross-entropy), supported languages (14), and performance metrics (WER, throughput). It highlights its #1 ranking on the HuggingFace Open ASR Leaderboard and provides comparative data against other ASR models. The post also outlines deployment options including local download, API access, and Model Vault, and mentions future integration with Cohere North.