Report · huggingface.co
Nvidia Nemotron 3.5 Transcribe: open ASR for 40 languages
huggingface.co
Source published October 28, 2026
Summary: 2 min read
53/100
Summary
NVIDIA released Nemotron 3.5 Transcribe, an open-weight speech model available through Hugging Face and ModelScope that combines transcription, speaker attribution, English translation, and audio question answering in one checkpoint. It transcribes 40 languages, translates 31 languages to or from English, supports speaker attribution in 13 languages, processes recordings up to one hour, and labels as many as eight speakers. Natural-language prompts can request outputs such as speaker-labeled transcripts, translations, meeting decisions, or specified formatting for technical vocabulary. Audio question answering is trained in English, and NVIDIA advises users to verify names, speaker changes, and answers against recordings. NVIDIA reports that the model ranks first among open-weight models on the Hugging Face English Open ASR Leaderboard with 5.33% average word error rate, and ties for first on Artificial Analysis’s open-weight ranking at 2.8% AA-WER while activating about 4 billion of its 30 billion parameters per token. Granary v2 supplies an open data-curation pipeline and public manifests covering 463,062 hours of speech across 48 languages. In an NVIDIA medical-transcription case study using Eka Care recordings, fine-tuning reduced held-out-speaker word error rate from 22.44% with the best prompt to 12.17%, a 45.8% relative reduction. NVIDIA said forthcoming quantized versions will support local deployment, while the current checkpoint can be served with vLLM and is designed for recorded rather than live-streaming audio.
Key Takeaways
- Nemotron 3.5 Transcribe is an open-weight model that performs multilingual transcription, speaker attribution, translation, and audio question answering from a single checkpoint.
- The model supports transcription in 40 languages, translation between English and 31 languages, speaker attribution in 13 languages, recordings up to one hour, and as many as eight speakers.
- NVIDIA reports a 5.33% average word error rate for Nemotron 3.5 Transcribe on the Hugging Face English Open ASR Leaderboard, the highest ranking among open-weight models there.
- Granary v2 provides public multilingual speech-data manifests covering 463,062 hours across 48 languages, plus a reusable curation pipeline and transcript-format labels for fine-tuning.
- In NVIDIA’s Eka Care medical ASR case study, fine-tuning reduced held-out-speaker word error rate by 45.8% relative to the best-prompt baseline.
What the discussion said
The thread quickly moved past the 40-language headline and into the practical question: is this a transcription model anyone would choose over the crowded open-ASR field? Several readers treated it as another capable but imperfect offline recognizer. One user’s hands-on experience found recurring mistakes even on ordinary vocabulary, while others said Parakeet and Qwen ASR currently deliver cleaner output. An early Japanese comparison was harsher still: Nemotron appeared slower, more resource-hungry, and less accurate than Parakeet, though the tester cautioned that quantization may have hurt it and planned to assess its built-in translation separately. The recorded-audio-only design was a clear disappointment for people who need dictation. A reader praised Nvidia’s streaming counterpart for handling German interleaved with English technical terms and product names, making the absence of live support here feel like an avoidable gap. Hardware accessibility also surfaced: prospective users wanted to know whether an old CPU-only machine could process audio at a tolerable overnight pace, rather than requiring an Nvidia GPU. Still, commenters identified one potentially distinctive capability: experimental near-field speaker transcription, an open speech model feature they had not seen elsewhere. For archival interviews with specialized vocabulary, however, the discussion offered no confidence that it beats a leading hosted OpenAI option.
Where opinion split
The sharp dispute is whether Nemotron 3.5 Transcribe is competitive with established ASR alternatives. Skeptics report a worse speed, memory, and error tradeoff than Parakeet or Qwen, especially for ordinary transcription; the more favorable case rests on unusual near-field speaker transcription and possibly embedded translation, features that could justify the heavier model for narrower workflows.
Coverage & discussions
Hacker News · huggingface.co
Community discussions
Community Sentiment
Positives
- Experimental near-field speaker transcription could fill a real open-source gap for recordings with identifiable nearby speakers, rather than merely chasing generic word-error benchmarks.
- Its embedded translation and broad language coverage may make the heavier offline pipeline worthwhile when transcription must cross language boundaries in one local workflow.
- Nvidia’s related streaming recognizer reportedly handles German speech mixed with English technical vocabulary and brand names unusually well, showing the family has practical multilingual dictation promise.
Concerns
- The model is explicitly aimed at recorded audio, leaving live dictation users without the capability that made Nvidia’s streaming variant compelling.
- Hands-on users report mistakes every few sentences on standard vocabulary, an unacceptable failure rate when transcripts must be trusted without constant correction.
- An initial comparison against Parakeet found it slower, heavier, and less accurate, suggesting its open release does not automatically make it the best local ASR choice.
- The apparent Nvidia-GPU expectation raises concern that CPU-only users may face painfully slow batch transcription, limiting the accessibility of a supposedly open model.
Related Articles
Voice AI in vogue: Nvidia releases new open transcription model
Oct 28, 2026
Cohere Transcribe: Speech Recognition
Mar 31, 2026
Voxtral Transcribe 2
Feb 4, 2026
Nvidia PersonaPlex 7B on Apple Silicon: Full-Duplex Speech-to-Speech in Swift
Mar 5, 2026
Pure C, CPU-only inference with Mistral Voxtral Realtime 4B speech to text model
Feb 10, 2026