ไทย
Bangkok and central Thailand at night from the International Space Station: amber city light spreading from the head of the Gulf of Thailand, green boat lights at sea, a thin green airglow line on the horizon.
Bangkok and the Gulf of Thailand at night, seen from the International Space Station, 6 May 2024.Image: NASA Licence

Tipjai Research

Measured in public.

Tipjai builds tools for Thai creators and the people who support them. Here we publish how well those tools work: how we measured, the numbers, and where we still fall short, written so you can follow it without being a researcher.

Read the latest work Tipjai Voice Lab
Source
ISS Expedition 71
Taken
6 May 2024
Points of light
22,030
Bangkok time
--:--:--

The field is drawn from a real photograph: Thailand at night from the International Space Station, 6 May 2024. Each point is one of its pixels. Image: NASA

Latest

Published work

  1. PaperHow 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.Evaluation, Speech, Thai language
  2. Model cardTipjai 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.Speech, Evaluation
  3. Research preview in trainingTipjai 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.Speech

Labs

Where the work happens

A long-exposure light painting of Lissajous figures: fine amber-white loops on pure black.
Lissajous figures, the curves two overlapping oscillations draw, painted with light.Dietmar Rabich / Wikimedia Commons / “Lissajous-Figur -- 2020 -- 7770” / CC BY-SA 4.0, resized Licence

Tipjai Voice Lab

Thai voices for live streams

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.

Heard wrong
8 of 30
Per alert
฿
Voices tested
in Thai

งานวิจัยที่ฟังได้

How we publish

Three things we do not bend

  • Blind by design

    Test sets are written apart and frozen before scoring. Nothing is tuned after seeing results, and whoever rates does not know whose work they are rating.

  • Numbers come from files

    Numbers on this site are read straight from the measurement output. None is typed by hand, so a page cannot drift from its results.

  • We show where we lose

    Losses are published with the wins, like the names Tipjai Voice 1 still gets wrong.

A studio meter bridge: ten backlit amber VU meters above rows of EQ modules lit violet.
Level meters in a studio. Every clip is matched in loudness before anyone listens.Photo: cottonbro studio / Pexels Licence