<!-- mobian-agent-page publisher="dailydev" canonical="https://daily.dev/posts/rumble-2-flash-an-open-speech-to-text-model-small-enough-for-a-phone-8lcq8vkco" -->

---
title: Rumble-2-Flash: an open speech-to-text model small...
description: Rumble-2-Flash is a free, open-weights speech-to-text model from Reson8 built for messy real-world audio. It has fewer than 1 billion parameters, small enough...
canonical: https://daily.dev/posts/rumble-2-flash-an-open-speech-to-text-model-small-enough-for-a-phone-8lcq8vkco
twitter:card: summary_large_image
twitter:site: @dailydotdev
og:type: website
og:site_name: daily.dev
og:title: Rumble-2-Flash: an open speech-to-text model small enough for a phone | daily.dev
og:description: Rumble-2-Flash is a free, open-weights speech-to-text model from Reson8 built for messy real-world audio. It has fewer than 1 billion parameters, small enough...
og:url: https://daily.dev/posts/rumble-2-flash-an-open-speech-to-text-model-small-enough-for-a-phone-8lcq8vkco
og:image: https://api.daily.dev/og/posts/8lCq8vkCO.png
og:image:alt: Rumble-2-Flash: an open speech-to-text model small enough for a phone
og:image:width: 1200
og:image:height: 630
og:locale: en
---

> ## Documentation Index
> Fetch the complete documentation index at: https://daily.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# Rumble-2-Flash: an open speech-to-text model small enough for a phone

**[Collections](https://daily.dev/sources/collections)** · 2 min read · 17 upvotes · 0 comments

## Summary

Rumble-2-Flash is a free, open-weights speech-to-text model from Reson8 built for messy real-world audio. It has fewer than 1 billion parameters, small enough to run on a phone, and supports over 40 languages. The weights are released under Apache 2.0 and work natively with Hugging Face Transformers and vLLM. It was trained on heavy accents, overlapping speech, phone calls and background noise, and @svpino points to its results on real phone calls as the part worth looking at.

## Content

Reson8, an Amsterdam startup, has released Rumble-2-Flash, an open speech-to-text model small enough to run on a phone. It has about 0.5B parameters, takes up roughly 1GB, and is licensed under Apache 2.0, so you can run it on your own machine and keep your audio local.

## What it's built for

Speech-to-text still breaks on real-world audio. Rumble-2-Flash was trained on the kind of recordings that usually trip these systems up: heavy accents, people talking over each other, bad phone lines, and street noise. It covers 40+ languages (one post puts the number at 45). Running on the device itself, it could transcribe calls and voice notes with no internet connection.

## Numbers

Reson8 says it competes with models 3 to 60 times its size.

- 6.96% error rate on the multilingual Open ASR Leaderboard
- 6.6% WER on English
- About half a second to process an hour of audio, with over 7,000 RTFx on full batches
- 7.86% WER on a 302-hour call-centre benchmark, close to NVIDIA's 30B Nemotron 3.5 Transcribe at 7.78%

That last result is the one I'd look at first. A model under 1B parameters landing within a hair of a 30B one on real phone calls is a big gap in size for almost no gap in accuracy.

## How it was made

Rumble-2-Flash was built through on-policy distillation from the larger Rumble-2 model, using the Qwen3-ASR architecture. It runs natively in Hugging Face Transformers and vLLM.

Speech recognition looks like one of the places where very small open models are closing the gap with frontier systems fastest, and this release fits that pattern.

## Questions this post answers

### What is Rumble-2-Flash and how many parameters does it have?

Rumble-2-Flash is a free, open-weights speech-to-text model from Reson8 with fewer than 1 billion parameters, small enough to run on a phone. It supports more than 40 languages, works natively with Hugging Face Transformers and vLLM, and is released under the Apache 2.0 license.

_Developers evaluating small on-device speech models can track releases like this one on daily.dev._

### Is there an open-source speech-to-text model that handles phone calls and background noise well?

Rumble-2-Flash was trained specifically on audio that usually trips up speech recognition, including heavy accents, overlapping speakers, phone calls, and background noise. Since its weights are open, it can run locally, letting teams handling sensitive call recordings keep audio off third-party servers.

_Those building voice pipelines for noisy audio can follow model comparisons like this on daily.dev._

## Community take

How the wider developer community reacted, aggregated from 3 discussions and 39 comments across x (as of 2026-11-19).

**TL;DR:** Developers are impressed that a sub-1B, Apache 2.0 speech model can run locally on a phone with no cloud round-trip, but many want proof it holds up on real-world messy audio like overlapping speakers, accents, and noisy calls rather than clean benchmarks.

**Sentiment:** 35% positive · 40% mixed · 25% skeptical

**The case for**

- On-device, local-only audio processing is seen as a real privacy and compliance win, especially for EU use cases.
- Running sub-1B models on a phone without API costs or round-trip latency is viewed as unlocking practical, always-available voice agents.
- Small footprint (around 1GB) makes it feasible to fit alongside other apps within typical phone RAM.

**The pushback**

- Skepticism that benchmark gains (e.g. beating 20x larger models) are shown on real messy audio rather than clean clips.
- No clarity on whether WER is reported per-language or blended across all 40+ languages, which could hide weak low-resource performance.
- Lack of detail on diarization or per-speaker WER for overlapping/crosstalk audio, a key real-world failure mode.
- Concerns that streaming latency in live production pipelines often undermines small architectures despite good offline benchmarks.

**By community**

- x (mixed): Excitement about an on-device, Apache 2.0 speech model is tempered by repeated requests for evidence on noisy, overlapping, and accented real-world audio rather than clean benchmark numbers.

**Hottest debate:** Whether the model's benchmark scores actually translate to real-world messy audio (noise, overlapping speakers, accents) or only hold on clean test clips.

**Open questions**

- Is the reported WER averaged across all 40+ languages or broken out per language?
- Does the model report WER specifically on overlapping/crosstalk audio, or only a blended score?
- Is diarization supported, or does it only produce a raw transcript without speaker separation?
- How does real-time factor hold up on a plain CPU in actual phone deployment?
- Does it chunk long audio similarly to Whisper's 30-second windows?

**Highlights**

> @omarsar0 Overlapping speech is the case everyone claims and nobody shows diarization WER on. What's the metric here, word error rate per speaker or just transcript-level WER ignoring who said what? Those two numbers tell very different stories on crosstalk.
> — [ukrroot on x](https://x.com/ukrroot/status/2123073748333187493)

> @omarsar0 The number that matters for phone agents isn't WER on noisy audio, it's which errors reach an action. "Fifteen" vs "fifty" on a booking is one token. Better models shrink that set, they don't empty it. Names, numbers and dates still need a read-back before anything writes.
> — [smperng on x](https://x.com/smperng/status/2123071942064529907)

> @svpino Sub-1B and 40 languages is the easy part, WER per language is what actually matters. Low-resource languages usually eat most of the error. Is the reported number averaged across all 40 or broken out per language?
> — [ukrroot on x](https://x.com/ukrroot/status/2123088848632852577)

> @rohanpaul_ai Beats models 20x its size on which audio though? Would love to know if that holds on the messy stuff, like overlapping voices on a bad phone line.
> — [JayAndAI on x](https://x.com/JayAndAI/status/2123088311757439152)

> @svpino under 1B is what makes the phone part real. 1B params x 2 bytes = 2 GB in bf16, x 0.5 byte = ~0.5 GB at 4 bit, so it sits easy next to everything else in 8 GB of ram. curious how it handles long audio, does it chunk at 30 s like whisper
> — [ArunSiddardha on x](https://x.com/ArunSiddardha/status/2123096810713649493)

**Source threads**

- [x](https://x.com/omarsar0/status/2123060794124386461) · 1 points · 20 comments
- [x](https://x.com/svpino/status/2123063550268874966) · 0 points · 12 comments
- [x](https://x.com/rohanpaul_ai/status/2123079032996827434) · 0 points · 7 comments

## Similar posts on daily.dev

- [Voice AI in vogue: Nvidia releases new open transcription model](https://daily.dev/posts/voice-ai-in-vogue-nvidia-releases-new-open-transcription-model-dotdulayg) · TechCrunch · 0 upvotes · 0 comments

---

Tags: [#speech-recognition](https://daily.dev/tags/speech-recognition), [#vllm](https://daily.dev/tags/vllm)

[View this post on daily.dev](https://daily.dev/posts/rumble-2-flash-an-open-speech-to-text-model-small-enough-for-a-phone-8lcq8vkco)

```json
{"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]}
{"@context":"https://schema.org","@type":"TechArticle","headline":"Rumble-2-Flash: an open speech-to-text model small enough for a phone","url":"https://daily.dev/posts/rumble-2-flash-an-open-speech-to-text-model-small-enough-for-a-phone-8lcq8vkco","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/rumble-2-flash-an-open-speech-to-text-model-small-enough-for-a-phone-8lcq8vkco"},"datePublished":"2026-11-18T14:40:05.538Z","dateModified":"2026-11-19T02:42:02.518Z","description":"Rumble-2-Flash is a free, open-weights speech-to-text model from Reson8 built for messy real-world audio. It has fewer than 1 billion parameters, small enough...","image":"https://pbs.twimg.com/media/HXajDFVbwAA6hTI.jpg","thumbnailUrl":"https://pbs.twimg.com/media/HXajDFVbwAA6hTI.jpg","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/rumble-2-flash-an-open-speech-to-text-model-small-enough-for-a-phone-8lcq8vkco","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":17},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"speech-recognition,vllm","timeRequired":"PT2M"}
{"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"Rumble-2-Flash: an open speech-to-text model small enough for a phone"}]}
{"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/rumble-2-flash-an-open-speech-to-text-model-small-enough-for-a-phone-8lcq8vkco#faq","mainEntity":[{"@type":"Question","name":"What is Rumble-2-Flash and how many parameters does it have?","acceptedAnswer":{"@type":"Answer","text":"Rumble-2-Flash is a free, open-weights speech-to-text model from Reson8 with fewer than 1 billion parameters, small enough to run on a phone. It supports more than 40 languages, works natively with Hugging Face Transformers and vLLM, and is released under the Apache 2.0 license. Developers evaluating small on-device speech models can track releases like this one on daily.dev."}},{"@type":"Question","name":"Is there an open-source speech-to-text model that handles phone calls and background noise well?","acceptedAnswer":{"@type":"Answer","text":"Rumble-2-Flash was trained specifically on audio that usually trips up speech recognition, including heavy accents, overlapping speakers, phone calls, and background noise. Since its weights are open, it can run locally, letting teams handling sensitive call recordings keep audio off third-party servers. Those building voice pipelines for noisy audio can follow model comparisons like this on daily.dev."}}]}
```

