Earlybird predicts spoken words before they end in 29M parameters
Original title: Earlybird: Predictive Streaming Speech Recognition in 29M Parameters
On LibriSpeech test-clean the median word is correct 90 ms early, and 69% of words are right before they end.
Researchers present Earlybird, a 29-million-parameter streaming speech recognizer that is 16.8 MB at 4 bits. It is trained on audio cut off at random points, including mid-word, and must emit every due word plus a guess of the next two. On LibriSpeech test-clean the median word is predicted correctly 90 ms before the speaker finishes it, while compared released streaming models lag 90 to 400 ms. Earlybird is earliest on all six timed sets and, on eight English Open ASR Leaderboard sets, beats Moonshine Tiny on every set and roughly ties same-size Whistle.
- The model has 29 million parameters and is 16.8 MB at 4 bits.
- Confidence for words still being spoken has ECE of 3.7 to 5.5.
- Timed sets cover audiobooks, TED talks, public speech, earnings calls and synthetic speech.
- On the same audiobooks, other models predict 3–25% of words correctly before they end.
- AI models
- EarlybirdMoonshine TinyWhistle
From the original article
1 min read at arxiv.orgStreaming speech recognizers output a word only after the speaker has finished it, typically 90-400 ms later. For a voice agent every millisecond counts: the sooner the words are outputted, the sooner the agent can start a search or plan its reply. We present Earlybird, a 29M-parameter streaming recognizer (16.
By Sahil Mahendrakar · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.