The speech recognition landscape is fragmenting in a way that could reshape how millions of people interact with AI. Rather than relying on a single general-purpose model like OpenAI's Whisper, companies are now building specialized speech-to-text systems tailored to specific languages, dialects, and real-world use cases. Two major releases this week underscore the trend: Cohere's new Arabic speech recognition model and NVIDIA's Audex, a unified audio-text system that handles both speech and general sound.

For Arabic speakers, the shift is particularly significant. Cohere released Cohere Transcribe Arabic, an open-source model that achieves a word error rate of 25.87 on the Hugging Face Arabic ASR Leaderboard, outperforming Meta's OmniASR by 2.45 points and OpenAI's Whisper Large V3 by 11 points. In head-to-head human evaluations with native Arabic speakers, the model was preferred over Whisper in 95.8% of tests. The breakthrough matters because Arabic presents unique challenges: more than 300 million speakers use roughly 30 recognized varieties shaped by distinct regional and cultural contexts, and everyday speech varies significantly across dialects, code-switching between Arabic and English, and domain-specific vocabulary.

What Makes These New Models Different From Whisper?

The core difference lies in specialization versus generalization. Whisper, released by OpenAI in 2022, was designed as a broad multilingual model that works reasonably well across many languages but doesn't excel in any single one. The new generation of models trades breadth for depth, focusing on specific languages or use cases where accuracy and cultural nuance matter most.

Cohere Transcribe Arabic, for instance, was trained extensively on data spanning Arabic dialects, professional language, code-switching, and varied acoustic conditions. The model preserves regional phrasing and dialect characteristics rather than converging toward Modern Standard Arabic, a critical distinction for users who speak colloquial varieties. Human evaluators assessed the model on three dimensions: overall accuracy, dialect faithfulness, and robustness to code-switching. Cohere Transcribe Arabic scored highest on all three compared with Whisper and the earlier Cohere Transcribe model.

NVIDIA's Audex takes a different approach by unifying audio and text in a single model. The system uses a 30-billion-parameter architecture that handles both speech recognition and general audio generation, meaning it can transcribe a German call and translate it to English, or generate sound effects from text descriptions like "birds chirping in a forest". The model avoids what researchers call the "text tax," a common problem where adding audio or vision capabilities causes text performance to drop. Audex maintains text intelligence comparable to its backbone model while adding audio capabilities.

How Do These Models Perform in Real-World Settings?

Performance benchmarks reveal where the new models excel and where gaps remain. On the Hugging Face Arabic ASR Leaderboard, Cohere Transcribe Arabic ranks first on four of six composite task sets. On the Casablanca dataset, which evaluates conversational Arabic across eight dialects, it improves on OmniASR by nearly six points. For production environments where speed matters, Cohere Transcribe Arabic achieves an RTFx (real-time factor multiple) score of 525 compared with 146 for Whisper Large V3, meaning it can process audio roughly 3.6 times faster than Whisper.

NVIDIA's Audex leads open-source models on speech recognition with an average word error rate of 6.82 on the OpenASR leaderboard, beating Step-Audio-R1.1-33B and Qwen3-Omni-30B-A3B-Thinking. However, the model shows mixed results on audio understanding tasks, leading on some benchmarks while trailing on others. The tradeoff reflects a deliberate design choice: Audex prioritizes simplicity and compatibility with standard machine learning infrastructure, running on Megatron-LM for training and vLLM for inference.

Steps to Integrate These Models Into Your Workflow

• Choose Your Use Case: If you need Arabic speech recognition with dialect preservation, Cohere Transcribe Arabic is available under the Apache 2.0 license on Hugging Face or through the Cohere API. For multilingual audio-to-text plus general sound generation, NVIDIA's Audex is released under a noncommercial license with model weights available for download.
• Set Up the Infrastructure: Cohere Transcribe Arabic is optimized for vLLM, a popular inference engine that handles high-volume speech workloads. NVIDIA's Audex also supports vLLM 0.20.0 and includes reference Docker containers for deployment, reducing setup complexity.
• Test on Your Data: Both models ship with quickstart implementations and example code. Cohere provides vLLM audio-QA scripts, while NVIDIA includes an interactive demo that computes token counts for any audio duration, helping you estimate costs and performance before full deployment.
• Monitor Real-Time Performance: For production use, track metrics like word error rate on your specific domain and RTFx scores under concurrent load. Cohere Transcribe Arabic's 525 RTFx score means it can handle high-throughput serving even when audio inputs vary in length.

The broader context matters here. A decade ago, building a voice-controlled system required expertise in speech recognition, natural language processing, and audio engineering. Today, developers can compose powerful capabilities by connecting APIs, as demonstrated by Collabora's recent experiment: a voice-controlled computer that transcribes speech, interprets commands, and executes actions in roughly 400 lines of Python. The shift from building models from scratch to composing them via APIs has collapsed the distance between "that would be a cool research project" and "that's a Saturday afternoon".

Why Does Language-Specific Development Matter for AI Adoption?

The emergence of specialized models addresses what researchers call the "frontier-language gap." While English dominates AI model development and evaluation, languages like Arabic remain underserved by state-of-the-art systems. This gap has real consequences: businesses in Arabic-speaking regions struggle to deploy accurate speech recognition, and millions of speakers are forced to use systems optimized for English phonetics and grammar.

Cohere framed its release as "a proud advance in the region's sovereign AI capabilities, bringing frontier performance to millions of Arabic-speakers". The emphasis on sovereignty reflects a growing recognition that AI infrastructure should reflect the linguistic and cultural diversity of its users, not just the convenience of developers building in English.

NVIDIA's Audex, meanwhile, demonstrates that unified models can preserve text intelligence while adding audio capabilities, a technical achievement that matters for enterprises deploying AI across multiple modalities. The model's ability to generate general audio beyond speech, combined with its strong speech recognition performance, positions it as a foundation for applications ranging from accessibility tools to voice assistants.

Both releases are available now. Cohere Transcribe Arabic is accessible through Hugging Face, the Cohere API, or the Model Vault. NVIDIA's Audex is released under a noncommercial license with model weights and code available for download. The shift toward specialized, language-aware speech recognition is underway, and the gap between Whisper and the next generation of models is widening fast.

" }