Tipjai Voice Lab
Research you can hear.
We build the voice that reads tip alerts on stream, then measure it blind against every voice on the market. Every result can be heard.
The method
What we measure, in ten seconds
- A tip alert is typed
The 60 test alerts are written in advance; no voice has seen them.
- The voice reads it
Supporter name, amount and message, as on a live stream.
- Machines listen
2 Thai speech recognizers write down what they hear.
- Count what went wrong
The share of characters heard wrong. Lower is clearer.
xXLotusXx ส่งน้ำใจ 10 บาท ตาหน้าเอาใหม่ ยังไหวอยู่
โลตัสส่งน้ำใจสิบบาทตาหน้าเขาใหม่ยังไหวอยู่
The board
All 30 voices, one tap apart
Pick an alert, then tap any voice. They all read the same words at matched loudness, so you can compare them by ear.
Models
Our voices
Tipjai Voice 1
The first Thai tip-alert voice, built to run on the streamer's own computer.
- Heard wrong
- 8 of 30
- Per alert
- ฿
- Time per alert
- s
Tipjai Voice 2
The next voice works on naturalness, where the first is still behind the best.
- Naturalness, Voice 1
- /5
- Best voice that round
- /5
Research
Work from this lab
- How clearly do voices read Thai tip alerts? A blind test of 30 voices60 unseen alerts, two speech recognizers, 2,000 resamples: where Tipjai Voice 1 stands against 29 voices from 10 companies, and where it still loses.
- Tipjai Voice 1: the full model cardTipjai's first Thai tip-alert voice: what it is for, how it was measured, results, limits and responsible use. 6.0% heard wrong, 8th of 30, ฿0.
- Tipjai Voice 2: in trainingThe next voice aims at naturalness, where the first is still behind the best. Results appear here only once they are measured on a blind set.
