Paxa Labs
Sign in

Thai speech recognition, on a language written without spaces

ASR, speech recognition, speech to text, transcription. Four names for one job, and in Thai that job has three problems English does not have.

Automatic speech recognition turns a recording into text. Every engine does the same two things: it decides what was said, and it decides how to write it down. In English the second half is close to free, because the sound of a word and its spelling travel together and a space marks where one word stops.

Thai gives the second half all of the difficulty. The writing system runs words together, the way a number is said differs from the way it is printed, and an ordinary Thai sentence carries English inside it. A recognizer that handles English well can still hand back Thai nobody can search.

Try it nowRead the speech-to-text page

The words for it#

These names appear in product pages, papers, and API documentation, sometimes for the same thing and sometimes for parts of it.

Automatic speech recognition (ASR)

Also called speech recognition, การรู้จำเสียงพูด

The field, and the usual term in research writing. An ASR system takes audio and returns the words in it.
Speech to text

Also called STT, แปลงเสียงเป็นข้อความ

The product name for the same thing. An API called speech to text and one called ASR do the same job.
Transcription

Also called ถอดเสียง, ถอดความ

The output and the act of producing it. In Thai the accurate term is ถอดเสียง, and ถอดความ is widely used for the same request.
Word timestamps

Also called word spans, เวลาระดับคำ

A start and an end second on every word. Subtitles, read-along players, and jump-to-the-minute search are all built from them.
Diarization

Also called speaker turns, การแยกผู้พูด

Marking who spoke when. The transcript is split into turns numbered by first appearance, with no claim about anybody's identity.
Word error rate (WER)

Also called อัตราความผิดพลาดระดับคำ

It counts word substitutions, deletions, and insertions over the number of words spoken. Most speech benchmarks report it.
Character error rate (CER)

Also called อัตราความผิดพลาดระดับตัวอักษร

The same measure counted over characters. Thai results are usually reported this way, for the reason below.

What Thai does differently#

Three properties of the language, each of which changes what a recognizer can honestly be asked for.

A Thai word boundary is a decision, and every measure of a Thai transcript inherits it.

This is a fact about the writing system, and no engine can answer it by getting better.

  1. 01

    Words run together

    Thai puts spaces at phrase and sentence breaks, and usually not between words. Counting mistakes word by word means first agreeing where the words are, which is the segmenter's decision. Thai results are reported over characters for that reason.

    Sources:arXiv (2026)(opens in a new tab)SNLP (2007)(opens in a new tab)

  2. 02

    The sound and the spelling part ways

    Speech compresses. A number said aloud comes out in words, while the same figure is printed in digits, and a repeated word is said twice where the page prints ๆ. A transcript has to pick a convention and hold it.

    Sources:HLT-NAACL (2006)(opens in a new tab)

  3. 03

    English arrives mid-sentence

    A Thai speaker names a technical term, a brand, or a product in English without changing sentence. The transcript keeps both scripts in one line, and the words on either side of the switch still need their seconds.

    Sources:HASSS (2023)(opens in a new tab)

What a Thai transcript should carry#

Four things a recognizer can return, and what each one is worth once the text is in a product.

Paxa STT Lite returns all four. Recordings run up to 60 minutes, in Thai and English, and none of these options changes the charge. The model is a research preview, and its transcripts and its options may change between releases.

Questions#

Is ASR the same thing as speech to text?
Yes. ASR is the research name and speech to text is the product name. A page selling an ASR API and a page selling a speech-to-text API describe the same job.
Why is Thai accuracy reported in characters?
Word error rate counts mistakes word by word, and Thai does not mark where words end. Two segmenters split the same sentence differently and produce two different scores for one transcript. Characters are stable, which is why Thai papers report them.
Do you publish an accuracy figure?
No figure goes out without the method beside it. An error rate quoted without its recordings, its segmenter, and its convention is a number that cannot be checked or reproduced.
Does a Thai transcript keep English words?
It writes them as they were spoken. A Thai sentence that switches into English for a term keeps that term in English, and the word spans cover both scripts.
What does ถอดความ mean?
In careful Thai, ถอดเสียง is transcribing sound into text and ถอดความ is rendering meaning. Both are typed into search for the same request, and this is the page for either one.
Can I try it without writing code?
Sign in and upload a recording in the console. The free grant covers a first file, and the transcript comes back with its word spans and both conventions available.

Try it on your own recording

The free grant covers a first file, with word timestamps, speaker turns, and either convention, and no card.

Start freeRead the speech-to-text page