Speech Models

A 0.5B transcription model in 40+ languages, free weights: but its English error rate looks better in the press version than in the company's own blog

November 19, 20260 views阅读中文原文
A 0.5B transcription model in 40+ languages, free weights: but its English error rate looks better in the press version than in the company's own blog

TL;DR

On November 18, the Dutch company Reson8 released the weights of Rumble-2-Flash under Apache 2.0: a 500-million-parameter speech recognition model covering more than 40 languages, which the company says can trade blows with models 3 to 60 times its size. It is small, it is cheap, and the license really does allow commercial use. But the same headline numbers do not match between the official blog post and the version handed to reporters. Before you build on it, it is worth knowing which measurement each figure came from.

What this actually is

First, what it is not: Rumble-2-Flash is an automatic speech recognition (ASR) model. It turns audio into text. It does not listen and talk back, and it produces no spoken output.

The timeline is also easy to garble, so here it is in order:

  • October 14: the model repository appears on Hugging Face;
  • November 3: the model goes live in Reson8's API;
  • around November 10: Reson8 ships its flagship Rumble-2 model;
  • November 18: the actual news here — the Rumble-2-Flash weights become a free download.

So "release" in this case means open weights, not a first appearance. If it feels like you heard about this model last week, you did.

Why it matters

Three things together keep it from being just another small model with a nice leaderboard line.

The size is genuinely small. The model card records 508,631,552 parameters in BF16. At that scale it does not need a GPU cluster: single cards, edge boxes, and on-device use are all plausible.

The license is Apache 2.0. Not "research only", not "apply and wait". You can ship it inside a commercial product. On the transcription layer, licensing often decides whether you launch far more than a fraction of a point of accuracy does.

It aims outside English. Reson8's pitch to reporters is that leading speech models are trained mostly on English even though only about one in twenty people speaks it as a first language, and that languages such as Dutch, Swedish and German — spoken by millions — get treated as an afterthought. There is an obvious commercial motive behind that framing, but the underlying problem is real.

Reading the numbers properly

This is the part worth slowing down for. The same company, in two places, does not give quite the same figures.

Multilingual average error rate: 6.96% or 7.09%?

  • The official blog post says: on the multilingual track of Hugging Face's Open ASR Leaderboard, the model averages 6.96% error across Chinese, German, French, Italian, Spanish, Portuguese, Hindi, Dutch and English — better than Meta's omniASR-LLM 7B, a model 15 times larger, and ahead of Whisper large-v3.
  • The press version says: in the company's own runs of that leaderboard's multilingual tests, it averages 7.09% across seven languages including German, French and Dutch.

Both can be true. The language sets differ (nine versus seven) and the runs differ (official leaderboard versus company-run). But they are not the same measurement, and they should not be swapped for one another.

English error rate: 6.6% or "about 5%"?

  • Blog: 6.6% on the Open ASR Leaderboard's English track.
  • Press: roughly 5%.

That is a 1.6-point gap, which is not small on the WER scale. The blog number comes from a leaderboard others can check; the 5% is the company's own summary. For budgets and acceptance tests, use the first.

Low-resource languages: 17.6% against 39.82% — the strongest comparison here

Across ten lower-resource languages such as Maltese and Basque, Rumble-2-Flash averages 17.6% error, while Whisper large-v3 posts 39.82% on the same test. That is the widest gap in the whole set, and the clearest evidence that the model was not tuned for English alone.

Real call-centre audio: 7.86%, which is "level with" in the loose sense

On AppTek's call-centre benchmark — 302 hours of helpline calls across 14 English accents — Rumble-2-Flash scores 7.86%:

  • NVIDIA Nemotron 3.5 Transcribe (30B): 7.78%
  • Rumble-2-Flash (0.5B): 7.86%
  • Qwen3-ASR 1.7B: 7.4%
  • Qwen3-ASR 0.6B: 8.7%

The official wording is "on par with Nemotron 3.5 Transcribe". Strictly, it trails by 0.08 points, and it trails its own larger sibling Qwen3-ASR 1.7B by 0.46 points. Getting a 0.5B model into that neighbourhood is impressive. "On par" is rhetoric; "slightly behind" is the fact.

Worth noting why the sibling comparison is fair: the blog states that Rumble-2-Flash uses the Qwen3-ASR architecture for day-0 support. It is the same skeleton, distilled to a different size, so losing to a larger member of the same family is not a surprise.

Speed: 7,000 RTFx is a full-batch number

The blog says an hour of audio transcribes in about half a second, exceeding 7,000 RTFx on full batches. "Full batches" is the qualifier that matters: high throughput is not low per-request latency. The model targets transcription of prerecorded audio, not live calls. No streaming latency figures were published.

How it works

One technical detail is worth keeping. Rumble-2-Flash was distilled from the larger production model Rumble-2 using on-policy distillation: the student transcribes first, then the teacher grades that transcript token by token, giving its own probability for what should come next at each position. The student learns from its own mistakes instead of memorising the teacher's finished text, which is finer-grained than one score per clip.

The tricky part is that the two models use different tokenizers, because the student adopted Qwen3-ASR's vocabulary. A method called GOLD merges tokens where the two tokenizers split text differently, so both models are compared over the same text. That removes the usual requirement that teacher and student share a tokenizer, and it is a reusable trick for anyone doing distillation across model families.

How to use it

  • Run it locally. The weights are at Reson8/Rumble-2-Flash on Hugging Face, with native Transformers and vLLM support. No trust_remote_code, no forked runtime. That is less friction than many "open" releases that need patching first.
  • Control the output format. It exposes switches for digits versus spelled-out numbers, keeping or dropping fillers such as "um" and "uh", keeping or collapsing repeated words, and forcing a language or letting it detect one automatically. Subtitling and meeting notes want different settings.
  • Use the API. Reson8 also hosts it (live since November 3) on its own GPUs in the Netherlands and says it retains no audio or transcripts. It claims ISO 27001 certification and GDPR and EU AI Act compliance; treat vendor self-attestations as claims to verify.

Who it is for

  • Teams transcribing languages other than English, especially Dutch, German, Nordic languages and Indian languages that mainstream models handle poorly.
  • Teams embedding transcription in a product who do not want licence terms to block them (Apache 2.0).
  • Teams that need to keep audio on local or edge hardware.
  • Workloads on prerecorded audio where throughput matters and tens of milliseconds of latency do not.

When it is not a good fit

  • Custom vocabulary. The custom glossary, up to 50,000 terms without fine-tuning, remains exclusive to the flagship Rumble-2. For medical or legal audio full of proper nouns, that alone can disqualify the Flash model.
  • Streaming. No streaming latency was published; the positioning is fast transcription of recorded audio.
  • Precise per-call budgeting. The press figure for API pricing is "from €0.01 per hour of audio" (about $0.001). That is implausibly low — it would mean 1,000 hours for €10. Either a unit was reported wrong or it is a promotional floor. Check the vendor's own pricing page before modelling costs.
  • Best-in-class Chinese or Japanese. The model card lists zh, yue, ja and ko among its language tags, and Chinese is in the leaderboard's multilingual track, but no per-language results are published. Do not assume all 40+ languages are equally accurate.

Pitfalls to watch

  1. Reading 7,000 RTFx as "real time". It is full-batch throughput. Per-request end-to-end latency on a single long file is unpublished.
  2. Mixing the error rates. 6.96% is a nine-language leaderboard figure, 7.09% is a seven-language company run; 6.6% is the English leaderboard, 5% is the company's summary. State which one you are citing.
  3. Reading "on par with a 30B" as "equal to a 30B". 7.86% against 7.78% is a small gap at the same level, not a tie — and the 1.7B sibling does better.
  4. Forgetting it is a distilled derivative. It inherits Qwen3-ASR's architecture and vocabulary, which becomes a constraint when upstream changes or when you need deep customisation.
  5. Judging maturity by download counts. At the time of writing the Hugging Face page showed 453 downloads and 84 likes: hours old, with no community fine-tunes or field reports yet.

Your next step

First, skip the leaderboard and do something cheaper: pull 200 recordings from your own traffic, covering the languages you care about and your worst conditions (in-car, noisy, heavy accents). Run three systems on them — Rumble-2-Flash, whatever you use today, and Qwen3-ASR 1.7B, its direct upstream. Same audio, same scoring script, compare WER and cost per thousand hours.

Second, settle the custom-vocabulary question early. If you need it, the Flash line is closed to you today and you are looking at Rumble-2 through the API.

Sources

  • Official blog post (November 18, 2026): https://huggingface.co/blog/rumble-2-flash
  • Model card and parameters (Apache 2.0, 508,631,552 parameters): https://huggingface.co/Reson8/Rumble-2-Flash
  • Media report (November 18, 2026; company-run figures, API pricing, company background): https://siliconangle.com/2026/11/18/reson8-releases-rumble-2-flash-speech-model-as-a-free-open-weights-download/

This article is compiled from the official blog post, the model card and the media report above; where the two sets of figures disagree, both are noted in place.

Sources

Comments

0 published
Log inorsign up to leave a comment.

Loading comments...