Nvidia Releases Nemotron 3.5 Transcribe, an Open Model for 40 Languages of Speech-to-Text

Nvidia has released Nemotron 3.5 Transcribe, a free open-weight transcription model that handles audio in up to 40 languages, TechCrunch reported on October 28, 2026.

Open weights means the trained parameters are published for anyone to download and run, rather than being available only through a vendor's hosted API. The model also translates 31 languages to and from English. It has 30 billion parameters, the learned numerical values that encode its behavior, and can run on a single accelerator. It processes recordings of up to one hour and can distinguish up to eight speakers within a single file. It works with recorded audio, not live streams.

Nvidia also released version 2 of Granary, its open-source speech dataset. The company has not, in the reporting available, tied the two releases to a single training pipeline, so readers should treat them as separate artifacts that happen to ship together.

Nemotron 3.5 Transcribe sits within Nvidia's broader speech portfolio, which the company describes as a family of GPU-accelerated open models covering automatic speech recognition (ASR), text-to-speech and speech-to-speech. Its earlier Parakeet TDT 0.6B v2, a 600-million-parameter ASR model, was designed for high-quality English transcription when Nvidia published it in June 2025. The new model is considerably larger, and its 40-language coverage extends well beyond that English-focused predecessor.

The multilingual work predates this release. In 2022 the company reported an ASR model trained on a Kinyarwanda dataset that reached state-of-the-art performance, and a Telugu language effort that it framed as expanding its speech AI reach. Its NIM microservices for speech and translation, introduced in 2024, covered text translation across more than 30 languages. On September 30, 2026, Nvidia published guidance on fine-tuning Nemotron for Saudi Arabic dialects, describing a path to other languages. That guide shows the workflow an enterprise or research group would likely use to adapt the model to a dialect or low-resource language, rather than relying on the base checkpoint alone.

For practitioners, the specifications set the deployment profile. A 30-billion-parameter model that fits on one accelerator is a different proposition from a multi-node service. The one-hour ceiling per recording suggests batch pipelines for call-center archives, media libraries, meeting transcripts and research interviews. The recorded-only constraint rules out live captioning as a direct use case, so anyone building a streaming product will need a different model or an architecture built around this one.

The broader context here is the shift in how speech models are distributed. Open weights let teams run transcription inside their own infrastructure, which matters for data residency, latency control and cost predictability. They also move responsibility for evaluation, domain adaptation and error handling onto the adopter. A model covering 40 languages reports an aggregate capability. Word error rate, the share of words transcribed incorrectly, will vary by accent, acoustic environment, and code-switched conversation where speakers mix languages mid-sentence.

In my view, the detail most likely to matter over time is the pairing of open weights with a dataset release. Speech systems are only as good as the data behind them, and a public dataset gives other teams a way to test claims, retrain for underserved languages and compare results on common ground. Whether that openness produces the gains its proponents expect will depend on how the community uses it.

The release also fits a pattern in which speech recognition is becoming a commodity component inside larger systems, with transcription feeding summarization, search and agentic workflows. For developers, a free model covering this many languages lowers the cost of experimenting with those pipelines, which is where much of the practical value is likely to appear first.