Skip to content

Rumble 2 Flash model competes with speech-recognition giants and lowers transcription costs for regional teams

Share
Rumble 2 Flash model competes with speech-recognition giants and lowers transcription costs for regional teams

Listen to this article

Read by Anchor

Reson8 launched its new speech model “Rumble 2 Flash” with 0.5 billion parameters under an open-source Apache 2.0 licence, breaking the traditional size-vs-performance equation that favours models many times larger. Our evaluation found that its inference efficiency rivals systems far larger, ranging from three to sixty times better, opening a path for regional development teams and contact-center hubs to lower the cost of locally transcribing and analysing audio without needing massive cloud compute.

The model was built on the “Qwen 3 ASR” architecture to run out-of-the-box without custom code.This design gives it native, direct support in the Transformers and FL-LM libraries, avoiding the need to run external code remotely. To reach this small size without sacrificing accuracy, the developers distilled their larger production model “Rumble 2” using interactive policy distillation; the student model transcribes audio itself while the teacher model corrects the outputs and predicts the next token probabilities at each position, providing a token-level evaluation instead of relying on an overall segment score.

To handle vocabulary differences and tokenisation mismatches between the two models, the team applied a technique they call “Gold”. The technique merges tokens when the tokenisation differs between the vocabularies, allowing a precise comparison on the same text by matching shared tokens directly and sorting the probabilities of the remaining tokens, which enabled accurate knowledge transfer from the larger teacher model to the smaller architecture.

In terms of speed and throughput, the processing rate exceeds 7 000 × real-time when running full-batch data on the Hugg­ing Face benchmark board, effectively transcribing an hour of audio in roughly half a second.This small model’s multilingual speech-recognition accuracy surpasses that of much larger giant models.Across a multilingual track covering 40 languages, it achieved a 6.96 % word-error rate on nine major languages, including Chinese, German, French, Italian, Spanish, Portuguese, Hindi, Dutch and English, outperforming Meta’s “Omni-ASR LLM” with 7 billion parameters and also exceeding the performance of the “Whisper Large v3” model.

Real-world phone-call testing created the most pronounced gap in field evaluation; on the “AB-Tech” metric used by contact-center teams, which includes 302 hours of support calls across 14 English accents, the model recorded a 7.86 % word-error rate. This places it roughly on par with Nvidia’s “NeMo-Tron 3.5 Transcribe” model of 30 billion parameters, which logged 7.78 %, while clearly beating the “Qwen 3 ASR” model of 0.6 billion parameters and offering fine-grained control over outputs such as spelling out numbers or using symbols, detecting filler words and literal repetitions.

What this means for the region

The arrival of an open-source model of this modest size and advanced performance represents a genuine turning point for technology leaders in Gulf banks, telecom operators and government service centres in Egypt and the Levant. Data-protection and banking-confidentiality regulations in the region impose strict limits on sending call recordings to external cloud APIs, while running massive models like “NeMo-Tron” locally requires costly servers and processing arrays that are hard to justify budget-wise.

The Apache 2.0 licence, together with immediate integration in FL-LM, lets regional engineering teams deploy speech-transcription systems on modest internal servers, cutting infrastructure bills by a record margin while keeping response times near-instantaneous.The operational bet is no longer tied to reserving large processing cards for transcription, but to the ability of local teams to embed half-billion-parameter models directly into customer-service pipelines.

Don't miss the next story

Subscribe for updates