ASR Watch · Edition 3 · 4 Nov 2026

The open transcriber grows up

What changed in speech recognition from 28 October to 4 November 2026: an open model that scores like the closed ones, a new #1, and a week of launches about everything around the words.

Prepared for Swaroop Narasimha · Sources on the last slide

Note This was researched and produced with AI. I have read it and decided it is worth publishing.

Space, Page Down or arrow keys to move. Or use the links at the bottom.

The short version

Three things from this week

2.8%Nemotron 3.5 Transcribe on Artificial Analysis, tied for best open weights.6 3.61Zoom Scribe V2.1 Pro, new #1 on the HF Open ASR English board (9-set mean).8
  • NVIDIA released Nemotron 3.5 Transcribe. Open weights, 40 languages, speaker labels, translation and prompts in one checkpoint.1,2
  • Zoom took #1 on Hugging Face. Ahead of Otter, last week's leader.7,8
  • The work moved around the words. ElevenLabs Cue for word timing, Soniox v6 for echo, Deepgram and Pipecat for context.11,13,15

WER is word error rate, lower is better. "vendor" = self-reported. "X post" = one post, not checked. Only news since 28 Oct. Not a quiet week.

Development 1 · NVIDIA, 28 Oct · bookmarked

Nemotron 3.5 Transcribe: one open checkpoint, many jobs

What it is

  • Mixture of experts, Mamba-2 plus Transformer. 34B total, about 4.3B active ("30B-A4B").1
  • Open weights, OpenMDW-1.1. BF16, FP8, NVFP4 and GGUF.1
  • Up to 1 hour of audio per call, 65,536-token context, one H100 80 GB.1
  • Released with Granary v2: 463,062 hours in 48 languages. vendor2

What it does

  • Transcription in 40 languages, translation in 31.1
  • Speaker-attributed transcripts in 13 languages, up to 8 speakers.1
  • Language ID, optional word timestamps, text prompts.1
  • Offline only, no streaming. A data-centre model, not a phone model.1

Launch post was in Swaroop's X bookmarks.4

Development 1 · The numbers

Close to the closed APIs, if you read the source

MeasureSourceResult
AA-WERArtificial Analysis (independent)2.8%, tied with Voxtral Small, 6x fewer active parameters5,6
AgentTalk (voice agent set)Artificial Analysis1.8% vs 2.3% for Voxtral Small5
Same model via DeepInfraArtificial Analysis3.4%: the host matters6
HF Open ASR EnglishBoard display / NVIDIA blog / tech report6.10 (#9) / 5.33 / 4.71 vendor7,2,3
Multilingual, long-formNVIDIA4.62% average, 10.98% long-form vendor3
SpeedNVIDIA100 hours of audio in under 3 min on one RTX PRO 6000 vendor2

The three English numbers differ because each averages different test sets. Always quote the source with the number.

The lesson to take

Prompt first, then fine-tune

NVIDIA's own results repeat last week's point about names, and add a second one about small domain training.

Names

An entity list cuts named-entity errors about six-fold on ContextASR-Bench. NVIDIA's test.3

Medical audio

WER 30.8% with no help, 23.8% with a prompt, 11.8% after fine-tuning. NVIDIA's case study.3

Held-out speakers

22.44% to 12.17%, a 45.8% relative cut, in about 14 GPU-hours on a GB300. NVIDIA's numbers.3

If audio has to stay on your own GPUs, this is the first open model to test. Bring a name list and a few hours of labelled audio.

Development 2 · Leaderboards

Zoom takes #1 on the Hugging Face English board

Model9 public sets (isitsota)Board display (fru.dev)Note
Zoom Scribe V2.1 Pro3.614.87new, around 31 Oct to 2 Nov8,7
Otter3.885.25last week's #18,7
p99lab audrey-0.75.67 (#3)new; $0.15/h files, $0.25/h streaming vendor10
Nemotron 3.5 Transcribe6.10 (#9)best new open model7

Zoom says it is the third time this year it has reclaimed #1. vendor9 The official dataset needs a login, so these are mirrors with different averaging. Last week's "new columns" question: Voice Arena Monsoon was added 28 Aug, so not new.

Development 3 · ElevenLabs, 28 Oct

Cue: word timing as its own product

179 msMean timing error on full recordings, against 283 ms to over 6,000 ms for open tools. vendor11 ~0.5 sTo align an hour of audio, against about 2 s for Qwen3-ForcedAligner. vendor11
  • A forced aligner. Audio plus text in, a time for each word out.11
  • Now does the word timing in all Scribe transcripts; sold through the Forced Alignment API.11,12
  • 57 ms on noisy English. Songs: 92.7% of words within 0.3 s, against 82.0%. 29 languages. vendor11
  • Why it matters: captions, chapters, clip cutting and tap-a-word-to-jump all depend on timing. No independent test yet.

Development 3 · Voice agents

Echo and context move into the recogniser

Soniox STT v6 · 3 Nov

  • The agent sends its own outgoing text to the recogniser.13,14
  • The model filters the agent's own voice and still hears the user cut in.13
  • One model, 60+ languages. No numbers published. vendor

Deepgram · 29 Oct

  • Acquiring Daily, the company behind open-source Pipecat; says Pipecat stays open and multi-vendor. vendor16
  • Real-Time Context Standard: three extra fields in the chat format for how the person sounded, what to do, and how to reply.15
  • Flux Multilingual: 10 languages, 31 more in early access, uses earlier turns. vendor15,17

For clean audio the words are mostly solved. Timing, echo, overlapping speakers and lost context are what break products, and that is where the launches went.

Small and on-device · India

Thirty-megabyte models for Indian languages

ModelWhatTheir result
Udgam Labs Shruti29.5 to 29.7 MB per language plus English, Apache-2.0, offline on phone, laptop or browser. Hindi (19 Oct), then Marathi, Tamil, Bengali, Telugu this week.18,19,20Hindi FLEURS 20.1 vs 28.9 for Whisper large-v3; English 21.4 vs 7.4 (worse). About 1 s per sentence, 80 MB on a 2021 Samsung phone. vendor18
Gnani Prisma v3.010 Indian languages incl. code-mixed, trained on phone calls.21,2216.93% macro WER (Vistaar) vs 19.98% Scribe v2, 20.51% Azure, 26.57% Nova-3. vendor21

Shruti keeps English words in English inside the native script, the way people type mixed speech. Its card is clear about where it loses.18

Smaller items

Apple, a tiny cleaner, and two papers

AnySpeechRecognizer · 29 OctSwift package on Apple's speech API; swap in Parakeet or Qwen3-ASR on Core ML, speaker labels for up to 4. Author's Mac: 1.55% test-clean (Parakeet v3) vs 2.12% for Apple's transcriber; an hour in 20 to 40 s. Not tested on iPhone.23,24
NVIDIA RE-USE · bookmark9.6M-parameter cleaner for noise, echo, clipping, phone audio. Released 22 Oct. Character accuracy after cleanup 89.17 vs 87.28 for a 500M rival. vendor25 M1 port at about 0.23 s per second of audio. X post26
Fluid Speech ProWeights are up: 25 languages, CC BY 4.0, "clean mode" drops fillers. Card "coming soon", no numbers yet.27
ufakzeka-kulak31.6M Turkish model from scratch: 9.21% FLEURS Turkish; beats Whisper small, not large-v3 (6.04%).28
ASR on an ESP32-P465.7M streaming model squeezed to 26.65 MB: 4.34% test-clean vs 3.94% for the original.29

Leaderboards · Artificial Analysis · as of 4 Nov

The top five did not move

ModelWhoNoteAA-WER
Fun-Realtime-ASR-previewAlibaba1.7
MAI-Transcribe-2Microsoft2.0
Voxtral Small Transcribe 3Mistral2.1
Scribe v2ElevenLabs2.2
Voxtral Mini Transcribe 3Mistral2.2
Nemotron 3.5 TranscribeNVIDIAnew, best open (tie)2.8

Artificial Analysis6: 15 of 62 scored models are open weights. Fastest is still Parakeet TDT 0.6B v3 on Together, now 613x real time (687x last week).

Quiet corners and context

Checked, nothing new in transcription

Quiet this week

  • OpenAI: no changes. whisper-1 and gpt-4o transcribe still shut down 26 Feb 2027.30
  • Google: Gemini 3.5 Transcribe still current.31
  • Microsoft: nothing since MAI-Transcribe-2-Streaming (1 Oct).32
  • Mistral: still no downloadable Voxtral 3 weights. Kyutai and AssemblyAI: nothing I could verify.

Context, not speech recognition

  • Speech-to-speech and TTS launches: ServiceNow Apriel Voice preview, Phonic-1, Hume.33,34
  • LiveKit Agents adds Reson8 and Voxtral 3; Pipecat 1.15.0 adds a tiny backchannel detector.33,34

What to try

Four things for this week

  1. Nemotron with and without a name list. A few hours of your own audio, a real list of catalogue names. Count names right each way.
  2. Against a small fine-tune. Check whether a Parakeet fine-tune from last week gets close enough at a fraction of the size.
  3. Test timing, not just words. On files with known boundaries, measure how far off Cue and an open aligner are.
  4. Clean first. Run RE-USE on your noisiest files and see whether WER moves.

Open questions and what I dropped

What I don't know yet

Open questions

  • Will independent long-form and diarization tests back up Nemotron's numbers?
  • A streaming or smaller Nemotron Transcribe?
  • What does Zoom V2.1 Pro cost, and how much does #1 change on your audio?
  • Any numbers for Soniox v6 echo handling, or an independent Cue test?
  • Will other frameworks adopt Deepgram's context standard?

Dropped as unverified

  • desertantlabs "Who" diarization speed (bookmark): video post only, no card or numbers to check.
  • An AssemblyAI "Universal-3 Pro" date from a search summary: no primary source.

Watching next week

Nemotron community tests, Zoom pricing, Cue tests, Soniox numbers, Pipecat after the deal, more Shruti languages, Fluid Speech Pro card.

Sources · checked 4 Nov 2026 · X bookmarks checked directly

Sources

  1. Hugging Face: NVIDIA Nemotron 3.5 Transcribe model card
  2. Hugging Face blog: Nemotron Transcribe and Granary v2
  3. NVIDIA: Nemotron 3.5 Transcribe technical report
  4. X post: @NVIDIAAI launch (bookmarked)
  5. X post: @ArtificialAnlys on Nemotron 3.5 Transcribe
  6. Artificial Analysis: speech to text board
  7. fru.dev mirror: HF Open ASR board
  8. isitsota: HF Open ASR snapshot
  9. Zoom: Scribe API
  10. Hugging Face: p99lab/audrey-0.7
  11. ElevenLabs blog: Cue aligner
  12. X post: @elevenlabs on Cue
  13. Soniox blog: STT v6
  14. Soniox docs: agent text input
  15. Deepgram blog: voice agents that keep the thread
  16. Deepgram blog: Daily and Pipecat join Deepgram
  17. Deepgram changelog
  18. Udgam Labs: Shruti
  19. GitHub: udgamlabs/shruti
  20. X post: @udgamlabs on Shruti
  21. Gnani: Prisma v3.0
  22. X post: @GnaniAi on Prisma v3.0
  23. Hugging Face blog: AnySpeechRecognizer
  24. GitHub: mattt/AnySpeechRecognizer
  25. Hugging Face: NVIDIA RE-USE
  26. X post: @TslShahir, RE-USE on M1 (bookmarked)
  27. Hugging Face: FluidVoice-Labs/fluid-speech-pro
  28. arXiv: ufakzeka-kulak, small Turkish ASR
  29. arXiv: streaming ASR on an ESP32-P4
  30. OpenAI API changelog
  31. Google: Gemini 3.5 Transcribe
  32. Microsoft AI: MAI-Transcribe-2
  33. hotwired.news Voice AI, week 44 2026
  34. hotwired.news Voice AI, week 45 2026

Read the essay: The words are solved. Everything around them is not.. Audio deep dive: Vaultcast, Speech, Real Systems #3.