# A 17 MB speech recogniser that often names words before they end

A single-author preprint retrains a small open streaming model so its median word appears about 90 ms before the speaker finishes it, each word tagged with a usable confidence. The cost is more words taken back, and it is less accurate than larger models.

- Category: Vision & multimodal
- Readiness level: III (Open source) on Science Underway's four-step scale, as proposed
- Published: 2026-11-12
- Editor: Stefan Perdal
- URL: https://scienceunderway.com/articles/a-17-mb-speech-recogniser-that-often-names-words-before-they-end

## The study

Earlybird: Predictive Streaming Speech Recognition in 29M Parameters (preprint, not peer reviewed). DOI 10.48550/arxiv.2611.14421, arXiv 2611.14421. https://arxiv.org/abs/2611.14421

## At a glance

- What it shows: On clean audiobook speech, Earlybird, a 29M-parameter streaming recogniser, gets the median word right 90 ms before it ends. Released models of similar or larger size get it 90 to 400 ms after. Its raw confidence scores closely match how often early words turn out right.
- What it does not show: It does not show that a real voice agent responds faster or better as a result. It does not test phones or embedded chips, and it is less accurate than Whisper base on most test sets.
- Why it matters: Voice agents wait for the recogniser before they can search or plan. Earlier words with trustworthy confidence could give back a few hundred milliseconds per turn, from a model small enough for a device.

## What the study shows

Streaming speech recognisers, which turn audio into text as it arrives, normally output a word only after the speaker has finished it. This preprint, which has not been peer reviewed, describes Earlybird, a model that tries to output each word while it is still being spoken, plus a guess at the next two words.

The author starts from Moonshine streaming Tiny, an open model released under the MIT licence, and trims its vocabulary to bring it from 34M to 29M parameters. It is then retrained on recordings cut off at random 80 ms points. At each cut, the target text includes every word that is 'due', set at roughly 300 ms before the word ends, written out in full even if only its opening sounds have been heard. Because each later target only extends the earlier one, the model is never trained to write text it later deletes. Quantisation, storing weights in fewer bits, shrinks the file to 16.8 MB at 4 bits.

Each word carries a confidence: the product of the probabilities the model gave its text pieces. No extra calibration step is applied.

The author fed every model audio in 80 ms chunks and logged when each word first appeared correctly and then stayed unchanged. Processing time was left out, so the numbers reflect what each model chooses to output, not hardware speed. Comparisons were with Moonshine Tiny's official runtime, Whistle (a model of similar file size) and Whisper base re-run every 0.48 s.

The author reports that on LibriSpeech test-clean (67 utterances) Earlybird's median word arrives 90 ms before it ends, against 270 ms after for Moonshine Tiny, 400 ms after for Whistle and 90 ms after for Whisper base. It was earliest on all six timed sets, covering audiobooks, TED talks, public speech, earnings calls and synthetic speech, with medians between 46 and 90 ms early. On LibriSpeech clean, 68.8% of words were right before they ended, against 3.4%, 9.1% and 25.4% for the three baselines.

For words still being spoken, expected calibration error (ECE, the average gap between stated confidence and actual hit rate) was 3.7 to 5.5 points. Words shown at 90% confidence or above were right 91% to 96% of the time.

On the eight English sets of the Open ASR Leaderboard, run on about 200 clips per set, the 4-bit model's mean word error rate (WER) was 12.6%, against 14.4% for Moonshine Tiny's runtime, 12.2% for Whistle and 9.7% for Whisper base. The weights and a loader are public under the MIT licence.

## Our reading

In our reading, the useful idea is simple and cheap: change only the training targets, not the architecture, and a small recogniser learns to commit to words early. The gains in earliness are large and consistent across six test sets and three training seeds. The model also runs at about 39 ms per 80 ms step on one server CPU thread.

The confidence result matters as much as the timing. An early word is only worth acting on if the system knows when to trust it, and here the raw scores track accuracy well without any tuning. That gives a downstream agent a straightforward threshold to work with.

The trade-off is real. Earlybird takes back 1.29 words for every word spoken on clean audiobooks, against 0.16 for Whistle and 0.39 for Whisper base, and 2 to 6 on spontaneous or noisy speech. It is also less accurate than Whisper base, which is 2.5 times larger, on every set except clean audiobooks. We rate it level III, open source available: weights are released and it has public benchmark results, but no named organisation reports using it.

The rest of the article (what it means, limits, what to watch) is free for members: https://scienceunderway.com/articles/a-17-mb-speech-recogniser-that-often-names-words-before-they-end

## How this was made

An AI model read the study and wrote this article in the structure Science Underway uses. An editor reviewed it and approved it for publication. Check the original study before you rely on the details.

---
Source: Science Underway, https://scienceunderway.com/. Please link to the article when you cite it.
