Skip to content
AI.info

The Pulse

Reson8 releases open speech model for more than 40 languages

Reson8 has released the weights of Rumble-2-Flash, a 0.5-billion-parameter speech recognition model, under the Apache 2.0 license. The model was previously available through the company’s API for prerecorded audio.

Reson8 releases open speech model for more than 40 languages

Reson8 has released the weights of Rumble-2-Flash, a speech recognition model for more than 40 languages, under the Apache 2.0 license. The 0.5-billion-parameter model can be run using Transformers or vLLM without custom code, according to the company’s post on Hugging Face.

The model was already available through Reson8’s API, following a launch the company dated November 3. The earlier launch offered pay-as-you-go transcription starting at €0.01 per hour of audio.

What developers can use

The model can detect the language of audio automatically, or a developer can specify one. Developers can also control whether numbers appear as digits or words, whether fillers are retained and whether repeated words are transcribed verbatim.

The weights are available in the Reson8/Rumble-2-Flash repository on Hugging Face, and the model remains available through Reson8’s API. It uses the Qwen3-ASR architecture and accepts audio supplied as a local path, a URL, a NumPy array or a list of those inputs.

The company’s documentation describes the prerecorded API as returning a plain transcript without speaker labels, timestamps or vocabulary customisation. It lists Rumble-2-Flash as supported on the prerecorded endpoint, but not on the realtime or turns endpoints.

Reported test results

On the Open ASR Leaderboard multilingual track, Rumble-2-Flash averaged a 6.96% error rate across Chinese, German, French, Italian, Spanish, Portuguese, Hindi, Dutch and English, according to the release. The post says that result is better than Meta’s omniASR-LLM 7B, which it describes as 15 times larger, and ahead of Whisper large-v3.

On the English track, the model averaged a 6.6% word error rate, the post says. On AppTek’s call-centre benchmark, the test covered 302 hours of helpline calls with 14 English accents and used AppTek’s scoring script.

Reson8 reports a 7.86% word error rate on that call-centre test, compared with 7.78% for NVIDIA’s Nemotron 3.5 Transcribe, a 30-billion-parameter model. The same comparison puts Qwen3-ASR 0.6B at 8.7% and Qwen3-ASR 1.7B at 7.4%.

Throughput depends on how full the batches are. Reson8 also reports throughput of over 7,000 RTFx on larger Open ASR Leaderboard datasets where batches stay full, equivalent in its account to processing an hour of audio in about half a second.

How it was trained

Reson8 trained the smaller model using its larger production model, Rumble-2, through a method called on-policy distillation. The smaller model transcribes audio itself, and the larger model scores its output token by token; the student then learns to match the teacher’s probabilities.

The two models use different tokenizers, which split text into units for processing. A technique called GOLD aligns their output by merging tokens where the tokenizers divide text differently. Tokens shared by both vocabularies are compared directly, while the remaining tokens are compared by their sorted probabilities.

Sources

  1. Hugging Face
  2. Reson8
  3. Reson8

Companies in this story

Related stories

Explore

More articles