Preview Audrey 0.7: speech recognition for real calls
Join the waitlist
In preview

Audrey 0.7.

Speech recognition for real calls.

Audrey 0.7

Accurate where it counts

Speech to text you can act on. Phone calls, meetings and accented English, with the names and numbers intact.

Real-world audioPhone lines, crosstalk, noisy rooms and accented speakers.
Critical wordsNames, numbers and the other words your workflow depends on.
One API callSend audio over HTTP, get one transcript back. Long recordings supported.
Two-channel callsOne speaker per channel, with the transcript labelled by speaker.
Request access
Word error rate · Audrey 0.7Lower is better
Test setWhat it isAudrey 0.7
LibriSpeech cleanRead audiobooks1.18%
LibriSpeech otherHarder read speech2.78%
SPGISpeechFinancial calls2.29%
VoxPopuliParliament speeches, many accents2.37%
Earnings-22Earnings calls, global accents5.41%
GigaSpeechPodcasts and YouTube7.15%
AMIMeetings with crosstalk7.34%
Measured by p99lab with the Open ASR Leaderboard's public evaluation code. Word error rate, lower is better.
Results

Results

Word error rate on standard public test sets, from clean read speech to meetings and earnings calls. On the hosted API, Audrey runs about 60 times faster than real time. To compare with other systems, see the Open ASR Leaderboard.

Version
Audrey 0.7
Measure
Word error rate
Speed
About 60× real time, hosted API
Status
Preview

In 1952, Bell Labs built the first machine that could recognise speech. It understood spoken digits. They called it Audrey.