Nemotron 3.5 Transcribe BEATS Whisper? (I Tested It)
NVIDIA just released Nemotron 3.5 Transcribe, a free AI that turns a whole conversation into text and labels who said what. On the Open ASR Leaderboard it makes about a quarter fewer mistakes than OpenAI's Whisper. So I tested it myself on NVIDIA's free demo: a doctor's visit, a question about the audio, and my own language, Tamil. What you'll learn: → What makes it different: one model, many jobs, picked with one line of instructions → How good it is: the Open ASR Leaderboard and VoiceArena's test with several speakers → My four tests, recorded at real speed: speaker labels, a question, a Tamil transcript, a Tamil translation → The catch: the hardware it needs, the new smaller versions, and no live speech → How to try it free, and what it costs to use from code Honest trade-offs: Nemotron is #9 of 61 on the leaderboard; eight paid services score better. Its VoiceArena win is within the margin of error, and NVIDIA lists VoiceArena as a partner. Tamil is supported for transcription only, and my Tamil test clip comes from a dataset in its training data. One Hacker News tester found a compressed copy slower and less accurate than NVIDIA's older Parakeet. ⏱️ Timestamps 0:00 NVIDIA's free AI that labels speakers 0:36 One model, many jobs 1:18 How good is it? 2:11 I tested it: four tests 3:50 The catch 4:28 Try it free 4:45 What it means for you My tests (29 Oct 2026, NVIDIA's free demo, temperature 0) • Doctor's visit: PriMock57 day 1, consultation 1 (a real Babylon clinician, a Babylon employee acting as the patient; CC BY 4.0). Full 7 min 38 s visit: 2 speakers found, transcript in 8.7 s, 14.1 words wrong per 100 against PriMock57's human transcript (1,318 words, fillers removed). Whisper medium on my Mac: 15.3, with no speaker labels. Whisper large-v3 scored 28.7 on the same file, mostly from words it added: it got stuck repeating lines nobody said ("so, you've been working" five times), a known Whisper failure, so I left it out of the video as not a fair result. One recording is not a benchmark. • Question about one minute of the visit: "Which medicine did the doctor recommend, and how often should the patient take it?" Answer: "Paracetamol, two tablets up to four times a day for the first few days." It matches the doctor's words. • Tamil transcript: a 16 s FLEURS clip (CC BY 4.0; FLEURS is in the model's training sources). The first two phrases came back word for word, then the end was cut short: 21.8% of characters wrong, 4 of 17 words wrong. • Tamil to English: "We can begin to live a very happy life with an attitude." The real meaning: we can start living in harmony with the environment. The model card lists Tamil for transcription only. Numbers used (checked 29 Oct 2026) • Hugging Face Open ASR Leaderboard, English, average WER (lower is better), updated 28 Oct 2026: OtterAI speech-v3.1 4.55 (#1), Nemotron 3.5 Transcribe 5.33 (#9 of 61), Qwen3-ASR 1.7B 5.40 (#11), OpenAI gpt-transcribe 5.96 (#23), Whisper large-v3 7.16 (#41). Ranks 1 to 8 are paid services. • VoiceArena Diarization-Bench v1, 139 English recordings with 2 to 8 speakers, cpWER: Nemotron 3.5 Transcribe 32.43 (95% range 26.93 to 38.03), NVIDIA Parakeet + Nemotron 3 Diarization 34.18 (29.35 to 38.94), ElevenLabs Scribe v2 35.73. 14 systems. • Model card: 34B parameters in total, about 4.3B active per token; 40 languages; recordings up to one hour; offline only (not live streaming); licence OpenMDW-1.1, commercial use allowed. • Hardware: the full BF16 version (67.5 GB) is deployed on one 80 GB H100. Official FP8 (34.7 GB) and NVFP4 (21.6 GB) versions went public on 29 Oct 2026; NVIDIA's table lists NVFP4 on an RTX 5090 and FP8 on a 48 GB L40S. NVIDIA reports about 1% (FP8) and 3.6% (NVFP4) more errors, relative. • DeepInfra: $0.00200 per minute of audio, so $0.12 per hour. Sources NVIDIA model card: https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B NVIDIA on Hugging Face, launch post: https://huggingface.co/blog/nvidia/nemotron-transcribe-granary Free demo: https://huggingface.co/spaces/nvidia/nemotron-transcribe Open ASR Leaderboard: https://huggingface.co/spaces/hf-audio/open_asr_leaderboard VoiceArena Diarization-Bench: https://voicearena.com/diarization-bench DeepInfra: https://deepinfra.com/nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B Hacker News thread: https://news.ycombinator.com/item?id=50275128 PriMock57 (Babylon Health, CC BY 4.0): https://github.com/babylonhealth/primock57 FLEURS (Google, CC BY 4.0): https://huggingface.co/datasets/google/fleurs Which language should I test it on next? Tell me in the comments. Follow Mervin Praison for more AI news, explained simply. #Nemotron #NVIDIA #SpeechToText #Whisper #AINews
