ไทย

Model cardSpeech, Evaluation

Tipjai Voice 1: the full model card

Tipjai'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.

Published
Authors
Tipjai Research
PDF
Download

Listen to the summary0:00

Read in Thai by Tipjai Voice, our synthetic voice in preview. It may misread a word.

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

Summary

Heard wrong

8th of 30 voices

Voices clearly clearer

of 29

Per alert
฿

on the streamer's computer

Time per alert
s

laptop CPU

  • Heard wrong on 6.0% of characters across 60 tip alerts no voice had seen; 95% interval 4.6% to 7.5%. None of the 29 other voices was clearly clearer, and 22 were too close to call.
  • It loses most clearly on supporter names (26th of 30) and on naturalness (3.17 of 5 against 5.00 for the best voice, one listener).
  • Known issue: an amount of 555 baht is read as laughter (found 14 October 2026; to be fixed in the next version).

For creators It reads tip alerts about as clearly as the best paid voices on the market, at no cost, but it still gets some names wrong and it still sounds synthetic.

Overview

Version
Tipjai Voice 1 (internally jai-pulse-1), frozen 6 October 2026 at 14:43
What it does
Reads Thai tip alerts aloud on a live stream: supporter name, amount and message
Language
Thai, including names written in Latin letters
Status
Lab preview
Runs on
The streamer's own computer, no graphics card needed
Price
฿0 per alert

It builds on open-source work, with every licence honoured. No vendor's audio was ever used to train or tune it, and no real supporter message was used.

What we do not publish: how the voice is built, meaning its components, data, sizes and settings. Outcomes and methods are public so anyone can check them; the recipe stays in-house so the work cannot simply be copied.

Intended use and what it is not for

  • Intended: Reading tip alerts on a live stream
  • Intended: Thai and Latin-script names, numbers, amounts, 555
  • Not for: Imitating a real person
  • Not for: Long reading, or other languages
  • Not for: Passing synthetic speech off as a person

How it was measured

Measured on 60 tip alerts written apart and frozen before scoring, against 29 voices taken from each vendor's own service. 2 Thai speech recognizers listened to every clip; the score is the share of characters heard wrong; intervals come from 2,000 resamples. Naturalness comes from one listener who did not know whose voice was whose.

The full method is in the paper, with the test set to download

Results

Bold is the best value in its row, whoever owns it
Tipjai Voice 1oursGemini 3.8 Flash TTS Koreclearest on the boardGoogle Standard-Acheapest paid voice
Heard wrong6.0%5.3%6.3%
95% interval4.6% to 7.5%3.4% to 7.7%3.9% to 8.8%
Rank range1 to 221 to 143 to 18
Name heard right (1 / 2)29 / 3041 / 4339 / 40
Whole alert right16 / 6025 / 6024 / 60
Naturalness (1 to 5)3.17--
Per 1,000 alerts฿0฿35฿6.7
Figure Tipjai Voice 1 is clearly clearer than 7 voices; none is clearly clearer than it; 22 are too close to call.
  1. Gemini 3.8 Flash TTS Kore+0.7 [−1.3, +2.7] too close to call
  2. ElevenLabs v4 Sarah+0.7 [−1.8, +2.8] too close to call
  3. ElevenLabs v4 Charlie+0.6 [−1.4, +2.4] too close to call
  4. Cartesia Sonic 3.6 Somchai+0.4 [−1.8, +2.4] too close to call
  5. ElevenLabs v4 Turbo Sarah+0.2 [−2.3, +2.4] too close to call
  6. MiniMax Speech 2.8 HD male 1+0.2 [−1.8, +2.0] too close to call
  7. Cartesia Sonic 3.6 Supannee+0.2 [−2.6, +2.5] too close to call
  8. Gemini 3.8 Flash Lite TTS Kore †−0.2 [−2.0, +1.6] too close to call
  9. Google Neural2-C−0.2 [−2.6, +1.9] too close to call
  10. Google Standard-A−0.2 [−2.6, +1.9] too close to call
  11. Cartesia Sonic 3.6 Kanya−0.3 [−3.2, +2.2] too close to call
  12. Grok Luna−0.3 [−3.6, +2.5] too close to call
  13. Microsoft MAI-Voice-2.1 Krit−0.3 [−3.1, +2.0] too close to call
  14. Grok Ara−0.3 [−3.9, +2.5] too close to call
  15. ElevenLabs v3 Conversational Sarah−0.8 [−3.5, +1.6] too close to call
  16. MiniMax Speech 2.8 HD female 1−0.8 [−4.0, +1.8] too close to call
  17. Microsoft MAI-Voice-2.1 Flash Harper−1.1 [−3.9, +1.6] too close to call
  18. Google Chirp 3 HD Kore−1.1 [−4.5, +1.7] too close to call
  19. Google Chirp 3 HD Charon−1.5 [−4.9, +1.3] too close to call
  20. Qwen Audio 3.0 TTS Flash, loongjohn−1.7 [−4.3, +0.5] too close to call
  21. Microsoft Premwadee−2.0 [−4.5, +0.3] too close to call
  22. ElevenLabs v3 Sarah−2.2 [−5.3, +0.5] too close to call
  23. Qwen Audio 3.0 TTS Plus, longanlingxin−2.5 [−5.0, −0.3] Tipjai Voice 1 clearly clearer
  24. Inworld TTS-2 Dennis−2.7 [−5.1, −0.5] Tipjai Voice 1 clearly clearer
  25. Microsoft Niwat−3.2 [−5.3, −1.1] Tipjai Voice 1 clearly clearer
  26. Microsoft MAI-Voice-2.1 Harper−5.4 [−11.4, −0.5] Tipjai Voice 1 clearly clearer
  27. Botnoi 311 (Northern dialect)−6.4 [−8.5, −4.3] Tipjai Voice 1 clearly clearer
  28. Inworld TTS-2 Flash Dennis−10.8 [−13.3, −8.4] Tipjai Voice 1 clearly clearer
  29. Fish Audio S2.1 Pro, library voice−20.7 [−24.1, −17.5] Tipjai Voice 1 clearly clearer

Each dot is the difference in errors (Tipjai Voice 1 minus that voice, in percentage points) on the same alerts; each line is its 95% interval.

Figure Naturalness: the next version's main job
  1. ElevenLabs v4 Sarah5.00
  2. ElevenLabs v4 Turbo Sarah5.00
  3. Qwen Audio 3.0 TTS Plus, longanlingxin3.50
  4. Tipjai Voice 13.17
  5. Microsoft Premwadee2.83
  6. Fish Audio S2.1 Pro, library voice1.50

1 listener, blind, 6 alerts. A small sample: a direction, not a ranking.

Speed and stability

On a laptop CPU it takes 1.38 s per alert (median of 20) and reads up to 76 alerts a minute in parallel. The same text gives the same audio every time.

Figure 400 stress inputs: no crash, no silence, no clipping
Inputs
400
Crashed, silent or clipped
0
Same text, same audio
Every time

The longest names and messages Tipjai accepts, emoji, links, other writing systems, numbers in every shape, invisible characters, HTML and stretched letters.

Safety and misuse

  • ImpersonationTipjai Voice 1 is one synthetic voice. It cannot be set to sound like a given person.
  • DisclosureIt is synthetic, not a recording of a person, and must not be used to pass as one.
  • Abusive textThe voice reads what it is given; Tipjai's alert system masks rude words before they reach it.
  • Report a problemA misreading or a misuse can be reported at tipjai.com/feedback.

Report a problem with the voice

Privacy

No real supporter message is in the data used to build or test this voice. Every test alert was written for testing.

Known limitations

  • NamesFewer names heard right than most voices (26th of 30).
  • NaturalnessIt still sounds synthetic, and naturalness was rated by one listener.
  • 555 bahtAn amount of 555 baht is read as laughter. Found 14 October 2026; the fix will be published with a fresh measurement.
  • The top group60 alerts cannot separate the top group.
  • RecognizersThe two disagree by up to 10 places, so the score estimates what a person hears.
  • LanguageThai only.

Versions and changelog

  1. Issue found: an amount of 555 baht is read as laughter. Added to the limitations.
  2. Naturalness round 2 added (1 listener, 6 alerts).
  3. A new way of reading was tried on blind set 3: worse, so not used.
  4. Same text, same audio every time; long messages read in full.
  5. jai-pulse-1 frozen at 14:43. Every number on this card was measured on it.
@misc{tipjai2026voice1card,
  title  = {Tipjai Voice 1: model card},
  author = {{Tipjai Research}},
  year   = {2026},
  month  = oct,
  url    = {https://research.tipjai.com/en/models/tipjai-voice-1/card/}
}