ASR Watch · Edition 1 · 21 Oct 2026

Small speech models got good

What moved in speech recognition from April to October 2026, and what it means for teams that build with small models.

A practitioner's report · Sources on the last slide

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

The short version

Three things to take away

~1 ptWER gap between the best open models and the best paid APIs on English, about 1 to 1.5 points.³ 26 Feb 2027OpenAI shuts down whisper-1 and the gpt-4o transcribe models in its API.⁷

Small open models are good enough. Models under 1B sit close to the 2B class and run much faster.^3,4
Whisper's API has an end date, and the replacements drop timestamps, subtitles and speaker labels.⁸
Speech on embedding models. A small adapter on EmbeddingGemma 2 gives very fast, cheap ASR for one language, at a real accuracy cost. A new 163 MB Parakeet fine-tune matches Parakeet on English.^34,35,37

WER is word error rate, lower is better. RTFx is times faster than real time. "vendor" = self-reported. "X post" = one post, not checked.

Leaderboard · HF Open ASR, English, as of Oct 20-21

Paid APIs lead, but small open models are close

Model	Type	Size	WER
Otter.ai Speech V3.1	API	n/a	3.88
Zoom Scribe V2 Pro	API	n/a	4.08
Azure Speech 06-2026	API	n/a	4.26
ElevenLabs Scribe v2	API	n/a	4.55
MOSS-Transcribe-Diarize	Open	n/a	5.17
Granite Speech 4.1	Open	2B	5.35
Edge0 ARK-ASR	Open	0.6B	5.53
Zipformer cr-ctc-transducer XL	Open	290M	5.56
Granite Speech 5.0 TurboCTC	Open	470M	5.78
Distil-Whisper large-v3.5	Open	n/a	6.18

Model

Type

Size

WER

Otter.ai Speech V3.1

API

n/a

3.88

Zoom Scribe V2 Pro

API

n/a

4.08

Azure Speech 06-2026

API

n/a

4.26

ElevenLabs Scribe v2

API

n/a

4.55

MOSS-Transcribe-Diarize

Open

n/a

5.17

Granite Speech 4.1

Open

2B

5.35

Edge0 ARK-ASR

Open

0.6B

5.53

Zipformer cr-ctc-transducer XL

Open

290M

5.56

Granite Speech 5.0 TurboCTC

Open

470M

5.78

Distil-Whisper large-v3.5

Open

n/a

6.18

Rows from the isitsota snapshot and the HF results dataset.^1,2,3 API rows have no speed figure. Which rows are vendor-submitted is not clear.

Development 1 · OpenAI

Whisper's API has an end date, and the replacements drop features

What changed

gpt-transcribe (files) and gpt-live-transcribe (realtime), released 28-29 Jul. Both take context, keywords and language hints.⁵
gpt-transcribe costs $0.0045 a minute, about $0.27 an hour.⁶
whisper-1 and the gpt-4o transcribe family shut down 26 Feb 2027.⁷
Live model: 19.70% error on Common Voice (22 languages) vs 20.33% for gpt-realtime-whisper-1. vendor⁵

Why it matters

No native SRT or VTT, no word timestamps, no diarization. OpenAI says to keep the old models for those until shutdown.⁸
If you use timestamps for chapters, aligned transcripts or clip search, this is a stack change, not a model swap.
Open Whisper weights are not affected.

Development 2 · Small open models

Sub-1B open models are production grade

The evidence

Leaderboard paper (86 systems, 12 sets): CTC and TDT decoders are fastest, transformer decoders most accurate.⁴
Cohere Transcribe 5.42 WER at 525 RTFx; Canary-Qwen 2.5B 5.63 at 418; Qwen3-ASR 1.7B 5.76 at 148.⁴
Parakeet TDT 0.6B v3: 25 European languages, 6.34% average WER. vendor^9,10 Default local engine in apps like Canonical's Myna.¹³
Parakeet Unified 0.6B: offline and streaming in one model, 160 ms minimum latency.¹¹ Nemotron 3.5 Streaming 0.6B: 40 locales, 80 ms to 1.12 s latency.¹²

Why it matters

For bulk transcription that feeds search or metadata extraction, a ~0.5B model on your own GPUs is within about a point of the 2B class and much faster. It is the same trade as with small text models.

Development 3 · Speech on an embedding model

How embedding-model ASR works

EmbeddingGemma 2 puts text, images, audio and video in one space, so a spoken word lands near the written word.³² Builders read words straight out of that space.

Word lookup (Haku)

The audio encoder stays frozen. A small reader maps each slice of audio to the nearest of ~229,000 English words; a tiny model adds punctuation. Only the reader is trained: under 2 h on 8 H100s, on openly licensed speech.^32,33

CTC head (Norwegian, Filipino)

norsk-asr uses the 12-layer audio conformer with only the top layer trained, a 6-layer conformer reader with CTC over 83 character units, and an optional 4-gram LM.³⁴ sea-asr-ph is the encoder plus a CTC head, streaming in 0.48 s chunks, with word timestamps and SRT/VTT.³⁵

And in reverse

The same backbone was used to plan speech for TTS: 79.5x real time on a base M3 Air, quality "not close to the best". X post³⁹ A later post claims 300x on an M5 Pro. X post¹⁵

Development 3 · The numbers

Very fast and cheap, clearly less accurate

Model	Language	Size	Speed	WER vs reference	License
Haku EG2 Transcribe	English	715 MB app models	288x, base M3 Air³²	~5% vs 2-3% for leaders; test set not named³²	Base Apache 2.0; free app
norsk-asr	Norwegian Bokmål	~850 MB with LM	1,000-1,600x H100; ~10x laptop CPU³⁴	FLEURS 10.0 vs 4.2 NB-Whisper large; NPSC 9.7 vs 16.5³⁴	Apache-2.0³⁴
sea-asr-ph	Filipino, Taglish	364M params	3,817x H100; 860x A10G³⁵	FLEURS fil 17.2 vs 10.6 Whisper large-v3; ties Whisper medium³⁵	Apache-2.0³⁵

Model

Language

Size

Speed

WER vs reference

License

Haku EG2 Transcribe

English

715 MB app models

288x, base M3 Air³²

~5% vs 2-3% for leaders; test set not named³²

Base Apache 2.0; free app

norsk-asr

Norwegian Bokmål

~850 MB with LM

1,000-1,600x H100; ~10x laptop CPU³⁴

FLEURS 10.0 vs 4.2 NB-Whisper large; NPSC 9.7 vs 16.5³⁴

Apache-2.0³⁴

sea-asr-ph

Filipino, Taglish

364M params

3,817x H100; 860x A10G³⁵

FLEURS fil 17.2 vs 10.6 Whisper large-v3; ties Whisper medium³⁵

Apache-2.0³⁵

Training data: 1,680 h of open Norwegian speech³⁴ and 1,914 h of CC-BY Filipino labels, released as a dataset.³⁵ All figures are the builders' own. norsk-asr's NPSC lead is mostly transcription style; counting only misheard words, NB-Whisper is 1.5 to 2.5x better.³⁴

Development 3 · What to make of it

Unverified, but the pattern is real

Not checked

"About $15 to train" for Norwegian and "10 to 40x faster than NB-Whisper" come from one X post. X post¹⁶
Haku's speed table mixes its own runs with others' benchmarks on different Macs; the page itself calls it "a rough guide".³²
None of these models is on the Open ASR Leaderboard yet.¹

Why it matters

The cost is tiny: a frozen backbone, a few hours of training, a few thousand hours of open data.^32,34,35
Error rates are 1.5 to 2.5x the best per language, so this fits search, draft captions and topic or entity extraction, not verbatim transcripts.
Google ships on-device products on the same backbone.¹⁸

New this week · Underdog

Underdog Bark 3.0: Parakeet accuracy in a 163 MB file

What it is

English ASR, a low-bit fine-tune of NVIDIA Parakeet TDT 0.6B v3 for on-device MLX. Not built on EmbeddingGemma. CC-BY 4.0.³⁷
4.94% average WER on the leaderboard's seven English sets, against 4.941% for Parakeet v3 (2.5 GB): a tie. vendor³⁷
Meetings (AMI): 7.82% vs 9.42% Parakeet and 13.63% Whisper large-v3. vendor³⁶
Family: Clean 180 MB drops "um"s (5.83%), Plus 180 MB 4.76%, Pro 1.25 GB 4.29%. vendor³⁶

What to watch

Speed claims differ: "1 min in under 0.25 s on a MacBook" in the launch post X post³⁸, about 1.5 s a minute (41x) on an M2 with stock mlx-audio on the model card.³⁷
Phonon-2 did the same trick at 164 MB and 5.21%; Bark edges it.³⁶
Self-scored with the leaderboard's method, not listed on it yet.³⁶
Page dated 15 Oct; announced on X 21 Oct.^36,38

Development 4 · Compression

Local transcription is a commodity

Fermion Phonon-2 · 164 MBParakeet v3 at ~2.1 bits. 1 h of audio in ~20 s on a MacBook Air; claims better than Whisper large. vendor¹⁹ Core ML port: 1 h in 5.6 s on an M5 Pro. X post²⁰

Cactus Whistle · 16.9 MBCPU, 7 languages, "mostly beats Whisper base". vendor¹⁴

Lokutor Oído · 13MRuns on a $5 ESP32-S3; 3.7% WER on LibriSpeech. vendor¹⁴

AuroraASR-1M · ~1 MB436x real time on one CPU thread, but 11.4% WER vs 5.6% for Whisper Tiny. X post²¹

QVAC TranscribePsy · 254M10 min in 1.65 s on iPhone 17 Pro Max. vendor License wording contradicts itself.¹³

Why it matters: transcription can happen where the audio is, so you may get text plus metadata instead of audio.

Development 5 · Prices

Batch is cents an hour, and every vendor ships a pair

Price per hour of audio

Microsoft MAI-Transcribe-2: $0.10, 60 languages, 5.2% WER on FLEURS. vendor²²
Gladia Solaria-3: from €0.16.¹³
OpenAI gpt-transcribe: ~$0.27.⁶
MAI-Transcribe-2-Streaming: $0.54, first partials in ~100 ms. vendor²³

Batch and live pairs

OpenAI: gpt-transcribe / gpt-live-transcribe⁵
Google: Gemini 3.5 Transcribe / -live²⁴
Microsoft: MAI-Transcribe-2 / -Streaming^22,23
AssemblyAI: Universal-3.5 Pro Async / Realtime²⁵
Mistral: Voxtral Mini Transcribe V2 / Realtime²⁶

Why it matters: at ten cents an hour, self-hosting only wins at real volume or when audio can't leave your systems.

Development 6 · Context

Give the model your entity list before you fine-tune

Context at inference

gpt-transcribe: prompts, keywords, language hints.⁵
Gemini 3.5 Transcribe: custom vocabulary, word timestamps, diarization.²⁴
MAI-Transcribe-2: keyword biasing.²²

Where people do fine-tune

ElevenLabs Scribe v2 Medical: ~35% lower WER on clinical audio. vendor²⁷
NatWest tuned Nemotron on UK accents: 20% fewer errors on its own test set. vendor¹³

Why it matters: in content understanding you usually know the names (people, products, places). Bias first, measure entity recall, then decide.

Development 7 · General LLMs

Multimodal LLMs still trail dedicated ASR

11.38Gemma 4 E4B WER on the leaderboard. The 12B scores 20.87.³ 7.4%of the time the Gemma 4 12B refused, in an independent test.²⁹

Gemma 4 E2B, E4B and 12B do ASR, in clips up to 30 s.²⁸
James Ding: E4B 7.6% mean WER on 7 sets without AMI; 12B hit 104% on AMI. "Don't swap it in for a dedicated ASR model."²⁹
Do this instead: a dedicated speech model in front, its text into the LLM.

Also worth knowing

Turn-taking, languages and licences

Turn-taking moves into STT

Deepgram Flux has built-in turn detection; Anam reports ~50% fewer barge-ins. vendor³⁰ ElevenLabs realtime Scribe added turn prediction on 12 Oct.¹³

Small languages

Covered by regional fine-tunes: Arabic dialects (Jais-ASR, Falcon-ASR), Japanese (Kotoba), plus Norwegian and Filipino.^13,16,17

Licence traps

Moonshine is non-commercial outside English. TranscribePsy's text contradicts itself. Cohere Transcribe is Apache-2.0 but needs the language set and has no timestamps.^13,31

What to try

Six experiments for this month

Baseline on your audio. 2 to 5 hours with good references; measure Parakeet v3, Bark 3.0 or Granite 5.0 470M for WER and cost per hour.
Score by the downstream task. Run your tagger on reference text and on each transcript. If tag F1 barely moves between 5 and 10 WER, use the cheaper model.
Bias before you fine-tune. Pass your entity list to gpt-transcribe or Gemini 3.5 Transcribe; measure entity recall.
Prototype an adapter on a frozen small backbone for one niche language or domain. Treat the X-post numbers as claims until you reproduce them.
Plan the OpenAI move. List everything using timestamps or diarization; it needs a new model before 26 Feb 2027.
Price it. $0.10 to $0.27 an hour from APIs against your own GPU cost per hour.

Open questions and what I dropped

What I don't know yet

Open questions

Which leaderboard rows are independent runs, and which are vendor submissions?
Does embedding-backbone ASR close the WER gap with more data?
How do the new OpenAI and Google models do on long audio against Parakeet TDT?
Is MAI-Transcribe-2 API only? It looks that way.
Will streaming benchmarks become as trusted as the HF leaderboard?

Dropped as unverified

Parakeet v3 at 6.32 WER / 3,330 RTFx in the paper: from a search summary, row not seen.
AssemblyAI "Universal-3.6 Pro Realtime": no primary source.
Nemotron 3.5 Streaming release date: sources disagree.
@reach_vb as a source to watch: not checked.

Sources · checked Oct 21, 2026