Sinhala speech-to-text, measured against Google.
Sinhala is a low-resource language, and the large providers price and prioritise accordingly. One has now declined to serve it outright.
The measurement
40 identical Sinhala clips. Both systems transcribed the same audio. Both scored against the same human reference transcripts, with the same scoring code.
| System | Word error rate | Character error rate |
|---|---|---|
| NOMOI Sinhala STT | 25.5% | 4.7% |
| Google Gemini (general model) | 42.7% | 16.4% |
NOMOI was more accurate on 27 of the 40 clips, tied on 6, behind on 7.
Google's newest transcription model does not support Sinhala at all. Released 26 August 2026 covering 85+ languages, it lists neither Sinhala nor Tamil, while Nepali, Bengali, Marathi, Telugu, Kannada, Malayalam, Gujarati, Punjabi, Burmese and Cantonese are all included. Sent Sinhala audio, it returns an empty result. The comparison above is therefore against Google's general-purpose model, which is the best Sinhala transcription Google currently offers.
What this number is not
The clips are clean read speech, from a public Sinhala corpus, held out so neither system had seen them. Read speech is an easier regime than a noisy newsroom, a phone line, or code-switched Sinhala and English.
So this is not a claim about your audio. We re-measure on your material before quoting you anything, and if our advantage does not hold there, we will say so and there is no pilot.
Send us 10 minutes of your audio
We return subtitles and a transcript within 48 hours. Free, no commitment. Judge the output the way you already judge captioning work: by reading it.
If it holds up: AED 900 (about USD 245) for up to 10 audio-hours, subtitle and text files, 72-hour turnaround. Before that quote we re-run both systems on a sample of your audio and show you both numbers.
Email assel@nomoi.ai