🔍 Intelligence Synthesis JP/EN
Ethics · Regulation · Technology — Pharma Practice Notes

AI Tech Flash ── 2026-10-29

Nvidia releases open-weight speech recognition model covering 40 languages

Nemotron 3.5 Transcribe turns recordings into text, translates, and accepts custom term lists

What changes

Nvidia has released "Nemotron 3.5 Transcribe," an AI model that turns recorded speech into text. According to BigGo Finance, it covers up to 40 languages. The model is "open-weight," meaning the numbers the model learned during training (its weights) can be downloaded by anyone.

According to South Korea's Financial News, developers can run the weights in their own environments and adapt the model for specific tasks as needed. The paper said the model can also be used to build multilingual speech recognition services, including for Korean.

Why it matters

BigGo Finance described the release as an effort to lower the barriers to entry in the speech AI market while broadening Nvidia's own developer ecosystem. It noted that speech AI models are launching in quick succession, including Mistral's Voxtral 3 earlier this month.

BigGo Finance also flagged caveats. Companies must operate and tune open-weight models themselves, so the cost structure and scope of responsibility differ from commercial API-based services. Whether performance holds up at benchmark levels in noisy real-world settings will help determine adoption, the outlet said.

What it is

Nemotron 3.5 Transcribe converts recorded audio files into text. According to BigGo Finance, it also translates among 31 languages, using English as the pivot language. It can distinguish up to eight speakers in a single recording.

Users can enter lists of technical terms or proper nouns for the model to recognize. The model has 30 billion parameters, the internal values that indicate a model's size. It processes only recorded files, not real-time streaming such as live calls or broadcasts.

Who and where

The model was built by Nvidia, which announced it on the 28th (local time), according to both outlets. Developers can download the weights and build them into their own services.

According to BigGo Finance, Nvidia said the model runs on a single chip, including the Jetson AGX Thor robot control chip, DGX Spark, and the Blackwell chip for data centers. Suggested uses include automatic meeting minutes and analysis of customer service call records.

Why it was built

According to the reports, fields such as medical care, finance, and law have tricky terminology that makes speech recognition hard. The term-list feature is meant to let companies build their own speech services in those fields.

Financial News said the release reflects Nvidia's strategy of using open-weight models to increase its influence not only in AI chips but also in AI development. BigGo Finance likewise said Nvidia is pushing to expand its touchpoints with developers.

How it works

A user feeds the model a recorded audio file. The model transcribes the speech and separates what each speaker says. For translation, it uses English as the reference language between 31 languages.

If a user supplies a list of technical terms or proper nouns, the model refers to it during recognition. According to BigGo Finance, Nvidia said combining the model with version 2 of its open-source speech dataset "Granary" and its "NeMo Speech" model can further improve recognition of industry terms in each language.

On performance, Nvidia claimed the model outperforms competitors on various benchmarks, including Hugging Face's Open ASR Leaderboard. ASR stands for automatic speech recognition, and a benchmark is a shared test used to compare models.

Key points

  1. Nvidia released Nemotron 3.5 Transcribe, which converts recordings in up to 40 languages into text, as open weights.
  2. It translates among 31 languages via English, distinguishes up to eight speakers, and accepts custom term lists.
  3. It handles only recorded files, not real-time streaming.

Conclusion

Nvidia has made a multilingual speech recognition model available for developers to download and adapt. According to BigGo Finance, whether its performance holds up in noisy real-world settings will be a key factor in adoption.

Quotes and sources

“It can distinguish up to eight speakers in a single recording, making it immediately usable for tasks such as automatic meeting minutes and analysis of customer service call records.”
── BigGo Finance, 2026-10-29
“One limitation, however, is that it processes only recorded files, not real-time streaming.”
── BigGo Finance, 2026-10-29
“Nvidia claimed the model outperforms competing models on various benchmarks, including Hugging Face's Open ASR Leaderboard.”
── BigGo Finance, 2026-10-29
“NVIDIA Corporation has made Transcribe available as open weights so developers can download the model weights, run them in their own environments, and adapt the model for specific tasks as needed.”
── Financial News (파이낸셜뉴스), 2026-10-29

Source articles

Written by AI (Claude) from the cited articles, then checked by Jev (TypeSafe's judgement model), a Gemini fact-check and a verbatim check of every quote.