Malaysian speech to text benchmark

Aisyah 1.0 Pro vs Google Gemini for Malaysian speech to text

Word error rates, where lower is better, for Aisyah 1.0 Pro, Google Gemini 2.5 Pro, Google Gemini 2.5 Flash and Google Gemini 3.6 Flash on Revolab's Malaysian speech benchmark: 1,899 clips across 12 categories.

Aisyah 1.0 Pro
Fewer mistakes overall, and on 6 of 12 kinds of audio. Biggest leads: short replies and phone calls.
Google Gemini 2.5 Pro
Fewer mistakes on 6 of 12: street interviews, FLEURS, singing, Common Voice, parliament and news. Also on noisy audio.

Revolab built this benchmark and Aisyah. Results as of 12 August 2026. How we tested

Word error rate on all 1,899 clipsMistakes per 100 words spoken. Shorter bar, fewer mistakes.
  • Aisyah 1.0 Pro5.79%
  • Google Gemini 2.5 Pro7.00%
  • Google Gemini 2.5 Flash10.76%
  • Google Gemini 3.6 Flash24.55%

Where Aisyah and Gemini 2.5 Pro each make fewer mistakes

Aisyah 1.0 ProGoogle Gemini 2.5 ProWord error rate. Further left, fewer mistakes.

Aisyah 1.0 Pro makes fewer mistakes on 6 of 12

  • Short repliesShort inputs, 155 clips
    Aisyah 1.0 Pro 3.28%Google Gemini 2.5 Pro 35.29%
  • Phone callsTelephony, 229 clips
    Aisyah 1.0 Pro 9.55%Google Gemini 2.5 Pro 19.80%
  • Animation150 clips
    Aisyah 1.0 Pro 5.66%Google Gemini 2.5 Pro 7.82%
  • Scripted readingRead speech, 154 clips
    Aisyah 1.0 Pro 1.30%Google Gemini 2.5 Pro 3.34%
  • PodcastsPodcast, 155 clips
    Aisyah 1.0 Pro 4.57%Google Gemini 2.5 Pro 6.09%
  • Drama156 clips
    Aisyah 1.0 Pro 5.91%Google Gemini 2.5 Pro 6.94%

Gemini 2.5 Pro makes fewer mistakes on 6 of 12

  • Street interviews143 clips
    Aisyah 1.0 Pro 14.24%Google Gemini 2.5 Pro 12.79%
  • FLEURSread-aloud research set, 153 clips
    Aisyah 1.0 Pro 3.42%Google Gemini 2.5 Pro 2.26%
  • Singing152 clips
    Aisyah 1.0 Pro 11.01%Google Gemini 2.5 Pro 10.15%
  • Common Voicevolunteer read-aloud, 141 clips
    Aisyah 1.0 Pro 3.74%Google Gemini 2.5 Pro 3.16%
  • Parliament155 clips
    Aisyah 1.0 Pro 3.79%Google Gemini 2.5 Pro 3.28%
  • News156 clips
    Aisyah 1.0 Pro 3.48%Google Gemini 2.5 Pro 3.24%

By background noise and clip set

The same 1,899 clips, split another way. The lower figure is in bold.

  • Clean audio1,364 clips
    Aisyah 1.0 Pro 4.70%Google Gemini 2.5 Pro 6.43%
  • Some background noise277 clips
    Aisyah 1.0 Pro 6.12%Google Gemini 2.5 Pro 6.65%
  • Noisy audio249 clips
    Aisyah 1.0 Pro 12.16%Google Gemini 2.5 Pro 11.13%
  • Published clips820 clips anyone can check
    Aisyah 1.0 Pro 4.89%Google Gemini 2.5 Pro 5.30%
  • Held-back clips1,079 clips no model has seen
    Aisyah 1.0 Pro 6.58%Google Gemini 2.5 Pro 8.49%

The other Google models on this page

  • Google Gemini 2.5 Flash (10.76% overall) makes fewer mistakes than Aisyah 1.0 Pro only on FLEURS (3.29% against 3.42%).
  • Google Gemini 3.6 Flash (24.55% overall) makes fewer mistakes than Aisyah 1.0 Pro only on FLEURS (2.79% against 3.42%).

Every figure for every model is in the full results below.

Hear the difference

Real clips from the published half of the benchmark. Play the audio, then read what each model returned, unedited.

Street interview about Malaysian films

Street interviews, 10.1 s
What was said
Filem Malaysia kurang sikit, tapi aa I suka ah macam cerita-cerita gangster semua. Filem apa you suka dan kenapa you suka? Macam KL Gangster, I tengok semua. Juvana.
Aisyah 1.0 Pro
Filem Malaysia kurang sikit tapi saya suka macam cerita-cerita gangster semua. Kenapa you suka dan kenapa? Macam cat gangster. Saya tengok semua Jovana.
Google Gemini 2.5 Pro
Filem Malaysia kurang sikit tapi aa I suka macam cerita-cerita gangster semua. Macam KL Gangster, I tengok semua, Juvana.
Google Gemini 2.5 Flash
filem Malaysia kurang sikit tapi ah I suka macam cerita-cerita gangster semua. Filem apa yang you suka dan kenapa suka? Macam Korean gangster. I tengok semua. Jovanna.
What to listen for

Google Gemini 2.5 Pro wrote "KL Gangster" and "Juvana" as in the reference, but its output has no text for the question "Filem apa you suka dan kenapa you suka?". Google Gemini 2.5 Flash wrote that question as "Filem apa yang you suka dan kenapa suka?", and Aisyah 1.0 Pro wrote "Kenapa you suka dan kenapa?". Aisyah 1.0 Pro also wrote "saya" where the reference has "I", and "cat gangster" for "KL Gangster". On this clip, Gemini 2.5 Flash's text is closer to the reference than Aisyah 1.0 Pro's.

Account number on a call

Phone calls, 2.5 s
What was said
102105
Aisyah 1.0 Pro
one, zero, two, one, zero, five
Google Gemini 2.5 Pro
10210
Google Gemini 2.5 Flash
one zero two one zero
What to listen for

The reference is "102105". Aisyah 1.0 Pro wrote the number as words, "one, zero, two, one, zero, five", which the benchmark scores the same as digits. Google Gemini 2.5 Pro returned "10210" and Google Gemini 2.5 Flash returned "one zero two one zero", both without the last digit of the reference.

Single-word Telephony clip

Phone calls, 1.2 s
What was said
Okay.
Aisyah 1.0 Pro
Okay
Google Gemini 2.5 Pro
[ 0m0s - 0m1s ] Okay.
Google Gemini 3.6 Flash
It looks like you didn't attach or link the audio, video, or text for the speech! Please provide the audio/video file, a link, or specify which famous speech you would like transcribed, and I will be happy to generate it for you.
What to listen for

Aisyah 1.0 Pro returned "Okay", matching the single-word reference. Google Gemini 2.5 Pro returned "Okay." with bracketed text in front of it, and Google Gemini 3.6 Flash returned a message asking for an audio or video file instead of a transcript.

Aisyah vs Google Gemini, answered.

Is Aisyah more accurate than Gemini 2.5 Pro for Malaysian speech?

On Revolab's Malaysian speech benchmark, Aisyah 1.0 Pro has the lower word error rate across all 1,899 clips: 5.79% against 7.00% for Google Gemini 2.5 Pro. Each has the lower WER in 6 of the 12 categories. Revolab built both the benchmark and Aisyah.

Which Gemini model is most accurate for Malaysian speech to text?

Of the Gemini models tested on Revolab's Malaysian speech benchmark, Google Gemini 2.5 Pro has the lowest word error rate across all 1,899 clips at 7.00%, followed by Google Gemini 2.5 Flash at 10.76% and Google Gemini 3.6 Flash at 24.55%. Aisyah 1.0 Pro is at 5.79%.

How accurate is Gemini at transcribing Malaysian phone calls?

In the Telephony category of Revolab's Malaysian speech benchmark, Google Gemini 2.5 Pro has a word error rate of 19.80%, Google Gemini 2.5 Flash 24.87% and Google Gemini 3.6 Flash 84.42%, where lower is better. Aisyah 1.0 Pro has 9.55%.

Is Gemini 2.5 Pro better than Aisyah on noisy audio?

On clips tagged noisy in Revolab's Malaysian speech benchmark, yes: Google Gemini 2.5 Pro has a word error rate of 11.13% and Aisyah 1.0 Pro has 12.16%. On clean clips, Aisyah 1.0 Pro has 4.70% and Gemini 2.5 Pro has 6.43%.

Check the numbers yourself.

Revolab built this benchmark and trained Aisyah, so read the results with that in mind. Every published clip and every model's output is open to check.

All the numbersEvery figure for all 4 models on this page, as tables

Overall

Word error rate for each model. Lower is better; the lowest in each column is highlighted.
ModelAll clips1,899 clipsPublished820 clipsHeld back1,079 clips
Aisyah 1.0 Pro5.79% (fewer mistakes)4.89% (fewer mistakes)6.58% (fewer mistakes)
Google Gemini 2.5 Pro7.00%5.30%8.49%
Google Gemini 2.5 Flash10.76%9.05%12.25%
Google Gemini 3.6 Flash24.55%15.01%32.87%

By kind of audio

Word error rate by kind of audio. Lower is better; the lowest in each row is highlighted.
Kind of audioAisyah 1.0 ProGemini 2.5 ProGemini 2.5 FlashGemini 3.6 Flash
Phone callsTelephony, 229 clips9.55% (fewer mistakes)19.80%24.87%84.42%
Short repliesShort inputs, 155 clips3.28% (fewer mistakes)35.29%81.79%753.63%
Scripted readingRead speech, 154 clips1.30% (fewer mistakes)3.34%7.04%10.75%
PodcastsPodcast, 155 clips4.57% (fewer mistakes)6.09%10.56%10.82%
Drama156 clips5.91% (fewer mistakes)6.94%9.52%15.82%
Animation150 clips5.66% (fewer mistakes)7.82%10.02%12.27%
News156 clips3.48%3.24% (fewer mistakes)3.93%5.08%
Parliament155 clips3.79%3.28% (fewer mistakes)6.47%7.24%
Street interviews143 clips14.24%12.79% (fewer mistakes)19.85%19.81%
Singing152 clips11.01%10.15% (fewer mistakes)20.30%28.14%
Common Voicevolunteer read-aloud, 141 clips3.74%3.16% (fewer mistakes)4.09%4.37%
FLEURSread-aloud research set, 153 clips3.42%2.26% (fewer mistakes)3.29%2.79%

By background noise

Word error rate by background noise. Lower is better; the lowest in each row is highlighted.
BackgroundAisyah 1.0 ProGemini 2.5 ProGemini 2.5 FlashGemini 3.6 Flash
Clean audio1,364 clips4.70% (fewer mistakes)6.43%10.07%27.78%
Some background noise277 clips6.12% (fewer mistakes)6.65%9.20%13.39%
Noisy audio249 clips12.16%11.13% (fewer mistakes)17.42%20.70%

Types of mistake

Types of mistake per 100 words spoken. Lower is better.
ModelWrong wordsper 100 wordsInvented wordsper 100 wordsDropped wordsper 100 words
Aisyah 1.0 Pro3.48% (fewer mistakes)0.62% (fewer mistakes)1.69% (fewer mistakes)
Google Gemini 2.5 Pro3.91%1.00%2.10%
Google Gemini 2.5 Flash5.50%2.48%2.78%
Google Gemini 3.6 Flash7.44%14.35%2.75%
How the test worksClips, scoring, noise tags and what is not compared
  • 1,899 Malaysian utterances in 12 kinds of audio, from phone calls and street interviews to parliament and singing. 820 are published and 1,079 are held back.
  • Word error rate counts every wrong, invented and dropped word and divides by the number of words actually said, across all clips. Lower is better.
  • Transcripts are normalised before scoring, so casing, punctuation and writing a number as digits or as words do not count as errors.
  • Clips are tagged for background noise: 1,364 clean, 277 moderate and 249 noisy.
  • Speed is not compared. For models served over an API, the time measured is mostly network latency rather than the model itself.
  • Results as of 12 August 2026, under the model names shown. Vendors update their models, so results can change.

Product and company names are trademarks of their respective owners and are used only to identify the models tested. Revolab is not affiliated with or endorsed by any of them.

Compare Aisyah with other models

See every model in one table

Test it on your own audio.

Every new account gets 50 free minutes of speech to text and 50 of text to speech. No card required.