Rumble-2-Flash

Parameters Languages License

Speech recognition model with native Hugging Face Transformers and vLLM support.

Usage (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.

Usage (vLLM)

uv pip install "vllm[audio]" openai
vllm serve Reson8/Rumble-2-Flash --max-model-len 4096
import base64
from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")
audio = base64.b64encode(open("audio.wav", "rb").read()).decode()

response = client.chat.completions.create(
    model="Reson8/Rumble-2-Flash",
    messages=[{
        "role": "user",
        "content": [{
            "type": "audio_url",
            "audio_url": {"url": f"data:audio/wav;base64,{audio}"},
        }],
    }],
    temperature=0,
    extra_body={"chat_template_kwargs": {"language": "English"}},
)
print(response.choices[0].message.content)

Transcription options

Pass these options through vLLM's chat_template_kwargs, or directly to processor.apply_chat_template in Transformers.

Option Default Effect
language unset Forces the language using its English name, such as "Dutch" or "English". If unset, output starts with language <Name><asr_text>.
digits true Numbers as digits; false writes them as words.
filler false true keeps fillers such as "um" and "uh".
verbatim true false collapses repeated words.

License

Licensed under the Apache License 2.0.

Open ASR Leaderboard results

Evaluated on the Open ASR Leaderboard (release 18-11-2026, v1.16.1), run on an H200 with vLLM. Compare against other open models on the multilingual tracks with this view (API models hidden).

English track

Average WER on the leaderboard: 6.60 (RTFx 6338), over the 7 public splits plus the aggregated Private (scripted) and Private (conversational) scores.

Split WER (%) RTFx
AMI 9.24 4245
Earnings22 6.45 4811
GigaSpeech 7.56 5978
SPGISpeech 3.08 7308
URGENT24 10.25 3314
Voice Arena Monsoon 3.56 3760
VoxPopuli 2.43 4222
Private (scripted) 3.71
Private (conversational) 13.14

Multilingual track

WER (%) per language and dataset; Chinese is CER. The last column is the leaderboard's average for the language.

Language FLEURS Common Voice MLS Voice Arena Monsoon AISHELL-1 Average
German 4.77 6.28 5.52
Spanish 2.82 4.33 4.04 3.73
French 5.95 10.01 5.66 7.21
Italian 2.53 4.84 7.90 5.09
Dutch 6.05 4.92 9.38 6.78
Portuguese 5.31 7.61 6.46
Hindi 14.11 14.11
Chinese (CER) 5.82 7.31 6.56
Downloads last month
453
Safetensors
Model size
0.5B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Spaces using Reson8/Rumble-2-Flash 3