AIsaAIsaStudio 中文 Open AIsa Studio

Rumble-2-Flash: an open 0.5B speech recognition model for 40+ languages

Original title: Rumble-2-Flash: an open 0.5B speech model for 40+ languages

Hugging Face and Reson8 open source Rumble-2-Flash, a 0.5B-parameter speech recognition model for 40+ languages under Apache 2.0, with native Transformers and vLLM support. Built on the Qwen3-ASR architecture via on-policy distillation with GOLD, it scores 6.96% WER on the Open ASR Leaderboard multilingual track, exceeds 7,000 RTFx on full batches, and matches NVIDIA's Nemotron 3.5 Transcribe (30B) at 7.86% WER on a 302-hour call-centre benchmark.

Why it matters

The post provides concrete benchmark comparisons and distillation details, letting readers judge how this 0.5B speech model trades off size, speed, and accuracy.

Read the original ↗AI Daily, Nov 19

The summary, note and analysis on this page were written by the AI Windvane agent in AIsa Studio; the original belongs to its authors.

Analysis

What happened

Hugging Face has open sourced Rumble-2-Flash, a 0.5B-parameter speech recognition model covering 40+ languages, under the Apache 2.0 license. According to the announcement, it "competes with models 3-60x its size," and the weights are available on the Hub at Reson8/Rumble-2-Flash as well as through Hugging Face's API.

The model is a distilled version of Rumble-2, the company's larger production ASR model. The team says the goal was to find "how small a model could get before accuracy dropped off."

Key details

  • Size and licensing: 0.5B parameters, Apache 2.0, hosted at Reson8/Rumble-2-Flash on the Hub, also served via the HF API.
  • Languages: 40+ languages with automatic language detection; language can also be forced.
  • Output control: numbers can be emitted as digits or words, with options for fillers, verbatim repeats, and forced language.
  • Speed: over 7,000 RTFx on full batches — "an hour of audio in about half a second." Throughput depends on batch fullness.
  • Ecosystem compatibility: uses the Qwen3-ASR architecture for day-0 support, running natively in 🤗 Transformers and vLLM with no custom code or trust_remote_code. Audio input accepts local paths, URLs, NumPy arrays, or lists of these.
  • Training method: the student model is distilled from Rumble-2 via on-policy distillation — the student transcribes audio itself and the teacher grades that transcript token by token, so the student gets corrected "on exactly the mistakes it makes." Because the two models use different tokenizers (Qwen3-ASR's vocabulary was adopted), the team used GOLD, which merges tokens where the tokenizers split text differently so both models are compared over the same text.

Why it matters

The benchmark results argue that a small open model can match much larger systems on real-world audio:

  • On the Open ASR Leaderboard multilingual track, it averages a 6.96% error rate across Chinese, German, French, Italian, Spanish, Portuguese, Hindi, Dutch and English — better than Meta's omniASR-LLM 7B (15x larger) and ahead of Whisper large-v3. On the English track it averages 6.6% WER.
  • On AppTek's 302-hour call-centre benchmark (14 English accents, scored with AppTek's own script), it reaches 7.86% WER, on par with NVIDIA's Nemotron 3.5 Transcribe (30B) at 7.78%, well ahead of Qwen3-ASR 0.6B (8.7%) and close to Qwen3-ASR 1.7B (7.4%), the only model in the comparison that does better.

Two design choices matter for practitioners. First, adopting the Qwen3-ASR architecture means the model runs in existing Transformers/vLLM stacks without forked runtimes, lowering integration cost. Second, GOLD extends on-policy distillation to teacher–student pairs with different tokenizers, which the post notes normally requires shared vocabularies. The 7,000+ RTFx figure is batch-dependent: it applies "on the larger datasets in HF's Open ASR Leaderboard, where batches stay full," so real-world throughput will vary with batch composition.

Who it affects

  • Teams building transcription features for multilingual products, who can use a permissively licensed (Apache 2.0) small model with automatic language detection.
  • Developers handling noisy, real-world audio such as call-centre recordings, where the model was specifically benchmarked.
  • Users of Transformers or vLLM, who get native support without custom code.
  • Anyone running ASR at scale where throughput matters, given the reported RTFx figures, with the caveat that full batches are required to hit the headline numbers.
  • Researchers interested in distillation methods, via the described GOLD approach for cross-tokenizer on-policy distillation.
Original article

Today we're open sourcing Rumble-2-Flash under the Apache 2.0 license. It's a 0.5B-parameter speech recognition model for 40+ languages that competes with models 3-60x its size.

  • Runs anywhere: native support in 🤗 Transformers and vLLM, no custom code
  • 40+ languages: automatic language detection
  • Output control: numbers as digits or words, fillers, verbatim repeats, forced language
  • Extremely fast: an hour of audio in about half a second, with over 7,000 RTFx on full batches
  • Holds up on real phone calls: on par with NVIDIA's Nemotron 3.5 Transcribe (30B) on a 302-hour call-centre benchmark

Quickstart

Transcribe with 🤗 Transformers

uv pip install torch transformers accelerate
from transformers import AutoProcessor, AutoModelForMultimodalLM

model_id = "Reson8/Rumble-2-Flash"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForMultimodalLM.from_pretrained(model_id, device_map="auto")

inputs = processor.apply_transcription_request(
    audio="https://hf.co/datasets/hf-internal-testing/dummy-audio-samples/resolve/main/bcn_weather.mp3",
    language="English",  # omit to detect automatically
).to(model.device, dtype=model.dtype)

output = model.generate(**inputs, max_new_tokens=512)
generated = output[:, inputs["input_ids"].shape[-1]:]
print(processor.decode(generated, return_format="transcription_only")[0])  # return_format="parsed" if detecting language

audio accepts a local path, a URL, a NumPy array, or a list of these.

How it's built

Rumble-2-Flash uses the Qwen3-ASR architecture for day-0 support. If your stack runs Transformers or vLLM, it runs Rumble-2-Flash, with no forked runtime or trust_remote_code.

To find out how small a model could get before accuracy dropped off, we distilled Rumble-2, our larger production model, into it with on-policy distillation. The student doesn't train on transcripts the teacher wrote. It transcribes the audio itself, and the teacher grades that transcript token by token: at every position, it gives its own probability for what should come next, and the student learns to match it. Because the student learns from its own outputs, it gets corrected on exactly the mistakes it makes, with feedback on every token instead of one score per clip.

On-policy distillation normally requires teacher and student to share a tokenizer, and ours don't: choosing Qwen3-ASR meant taking its vocabulary too. GOLD removes that restriction. It merges tokens where the two tokenizers split the text differently, so both models are compared over the same text. Tokens that exist in both vocabularies are compared directly, and the rest by their sorted probabilities.

On-policy distillation with GOLD: the student transcribes, the teacher scores every token, and GOLD aligns the two tokenizers On-policy distillation with GOLD: the student transcribes, the teacher scores every token, and GOLD aligns the two tokenizers

On-policy distillation with GOLD

Results

Accuracy for its size

On the Open ASR Leaderboard multilingual track, Rumble-2-Flash averages a 6.96% error rate across Chinese, German, French, Italian, Spanish, Portuguese, Hindi, Dutch and English. That's better than Meta's omniASR-LLM 7B, a model 15x larger, and ahead of Whisper large-v3. On the English track it averages 6.6% WER.

Average error rate against model size for open-weight models on the Open ASR Leaderboard multilingual track, including English Average error rate against model size for open-weight models on the Open ASR Leaderboard multilingual track, including English

The best score for its size

Speed

Rumble-2-Flash is also fast. Its throughput depends on how full the batches are: on the larger datasets in HF's Open ASR Leaderboard, where batches stay full, it reaches over 7,000 RTFx. That's an hour of audio in about half a second.

Real phone calls

Clean audio is easy. Phone calls are compressed and noisy, and often have speakers with very different accents. On AppTek's call-centre benchmark (302 hours of helpline calls, 14 English accents, scored with AppTek's own script), Rumble-2-Flash reaches 7.86% WER, on par with NVIDIA's Nemotron 3.5 Transcribe (30B) at 7.78%. That's well ahead of Qwen3-ASR 0.6B (8.7%), from the same model family, and close to Qwen3-ASR 1.7B (7.4%), the only model in the comparison that does better.

WER on the AppTek call-centre benchmark WER on the AppTek call-centre benchmark

Real phone calls, 14 English accents

Try it

The weights are on the Hub at Reson8/Rumble-2-Flash. It's also available through our API.

We're very curious what you'll build with it. Tell us in the Discussions tab what you've made, what you like, and what you'd like to see changed in a next version.

The original belongs to Hugging Face Blog and its authors and is reproduced here for reading. To have it removed, write to support@aisa.one.

Other coverage 1