# What is speech to text, and how does it work?

Speech to text turns spoken audio into written words. How speech recognition works, what a transcript contains, what people use it for, and what it costs.

Published 2026-11-17. Paxa Labs.

Speech to text, or STT, is software that listens to spoken audio and writes down the words. You give it a recording or a live voice, and it returns a transcript. The same technology is called transcription, speech recognition, and automatic speech recognition (ASR).

It is the reason a phone can take dictation, a video can carry captions made in minutes, and an hour of meeting can be searched for one sentence. Text to speech is the opposite job, and [the comparison guide](https://paxalabs.com/guides/text-to-speech-vs-speech-to-text) sets the two side by side.

## How does speech to text work?

Every system does three jobs in order, whatever it is built from.

1. Listen to the audio: The recording is cut into slices a fraction of a second long. Each slice is measured for the pitches it contains and how loud each one is, which gives the software a picture of the sound over time.
2. Match sounds to words: A neural network that has learned from many hours of speech paired with its transcripts reads that picture and predicts the words most likely to have produced it. Context settles the close calls. The words around a sound decide between two words that are pronounced alike.
3. Write the text: The words are written out with punctuation and sentence breaks. This step also decides how a number, a date, or a unit appears on the page.

Early systems recognized one speaker saying one word at a time from a short list. A modern system handles continuous speech from people it has never heard, which is why it works on a meeting recorded on a laptop.

## What is the difference between recorded and realtime speech to text?

Recorded transcription takes a finished file and returns the whole transcript at once. Paxa STT Lite works this way, on a file up to 120 minutes long. It suits meetings, interviews, lectures, and videos.

Realtime transcription listens to a live voice and returns text while the person is still in the conversation. Paxa STT Lite Realtime is streaming transcription with automatic turn detection. It reports when the speaker starts, when they stop, and what they said in that turn. Live captions and voice agents need this kind. [Realtime vs recorded transcription](https://paxalabs.com/guides/realtime-vs-recorded-transcription) covers how to choose.

## What do you get back from speech to text?

A transcript is the minimum. Most services can add more to it, and these are the parts worth knowing by name.

| Output | What it is | What it is for |
| --- | --- | --- |
| Transcript | The words that were spoken, as text | Reading, searching, quoting, and summarizing |
| Word timestamps | The second each word starts and ends | Jumping to a quote in the audio, and timing captions. See [the word timestamps guide](https://paxalabs.com/guides/word-level-timestamps-api) |
| Speaker numbers | The transcript split into turns, each tagged Speaker 0, Speaker 1, and so on | Telling an interviewer from a guest, or an agent from a caller |
| Caption file | An SRT or VTT file with every line timed | Uploading beside a video. See [the subtitle generator](https://paxalabs.com/subtitle-generator) |

Two settings change how the text reads. Paxa STT Lite writes Thai numbers, units, and dates either Spoken, in words as they were said, or Written, in digits the way a report prints them. [The numbers guide](https://paxalabs.com/guides/thai-transcript-numbers-spoken-or-written) shows both. Its style is either Verbatim, which keeps every filler and false start, or Clean, which drops them so the text reads more easily.

A vocabulary list of up to 50 names, products, and terms tells the model how to spell the words your recording is likely to contain.

> Paxa STT Lite is a research preview. Its transcripts and its options may change between releases.

## What do people use speech to text for?

| Job | What the transcript does | Read more |
| --- | --- | --- |
| Meetings | Becomes the draft of the minutes, with decisions and owners pulled from the text | [Meetings](https://paxalabs.com/use-cases/meetings) |
| Customer calls | Makes every call searchable for review, training, and complaints | [Call recordings](https://paxalabs.com/use-cases/call-recordings) |
| Video captions | Turns into a timed caption file for a clip, a course, or a broadcast | [Captions](https://paxalabs.com/use-cases/captions) |
| Interviews | Gives a researcher or a reporter exact quotes with the second each was said | [Interviews](https://paxalabs.com/use-cases/interviews) |
| Lectures and training | Becomes notes students can search and revise from | [Lectures](https://paxalabs.com/use-cases/lectures) |
| Voice notes | Turns a spoken reminder into text that can be pasted where it belongs | [Voice notes](https://paxalabs.com/use-cases/voice-notes) |
| Dictation | Writes a document while its author speaks | [Dictation](https://paxalabs.com/use-cases/dictation) |
| Live captions and voice agents | Supplies the words in each turn while the conversation is still going | [Live transcription](https://paxalabs.com/use-cases/live-transcription), [what an AI voice agent is](https://paxalabs.com/guides/what-is-an-ai-voice-agent) |

## How is speech to text accuracy measured?

The standard measure is word error rate, or WER. It counts the words a transcript got wrong, left out, or added, and divides by the number of words actually spoken. Lower is better. Languages written without spaces between words are usually scored by character error rate, CER, which counts characters the same way. [The ASR glossary](https://paxalabs.com/asr) defines both terms, and [the accuracy guide](https://paxalabs.com/guides/measuring-thai-transcription-accuracy) shows how to run the measurement on your own recordings.

A published score describes the recordings it was measured on. Clean studio speech scores far better than a phone call in a moving car, on every system. Test a service with a recording that sounds like yours before you rely on it.

## Why are some languages harder to transcribe?

- Words run together on the page. Thai and Chinese are written with no spaces between words. The software has to decide where each word ends before a timestamp or a caption line can sit on it.
- Tone changes the word. Thai has five tones and Mandarin four, and the same syllable on a different tone is a different word. In Mandarin, many words also share one sound, and only the sentence around them says which character to write.
- Languages mix inside one sentence. Thai speakers say a product name or a technical term in English and carry on in Thai. A good transcript keeps each word in the script it was spoken in, and [the code-switching guide](https://paxalabs.com/guides/transcribing-thai-english-code-switching) explains what that does to search.
- Numbers are spoken as words. Whether a transcript writes them as words or as digits is a setting, and the right answer depends on who reads it.

Paxa STT Lite is multilingual. It transcribes Thai, English, Mandarin, and other languages, and it detects the language when you do not name one.

## Does speech to text detect laughter, pauses, or who is speaking?

A transcript holds the words that were spoken. The Verbatim style keeps the fillers, false starts, and self-corrections a speaker made. A pause appears as the gap between one word's end time and the next word's start, and a caption file starts a new cue at a pause.

Speaker numbers tell the voices in one recording apart. They identify nobody. The transcript says Speaker 0 and Speaker 1, and a person reading it decides which one is the host.

## What does speech to text cost?

Speech to text is priced by the length of the audio. Realtime transcription has its own rate, billed on the audio received.

| Model | Credits | Money |
| --- | --- | --- |
| Paxa STT Lite | 8.33 per minute | $0.50 per hour |
| Paxa STT Lite Realtime | 12.5 per minute | $0.75 per hour |

A new account gets 100 free credits with no card, which cover about 12 minutes of recorded audio. They are for personal use and evaluation. Buy any plan or credit pack, and the account's transcripts are yours to use commercially, including those made before the purchase. [The Terms](https://paxalabs.com/terms#content-and-output) set this out, and [the transcription cost guide](https://paxalabs.com/guides/how-much-does-transcription-cost) compares these rates with paying a person to type.

## How do you try speech to text?

1. Open the Transcribe tool: Sign in and open [the Transcribe tool](https://paxalabs.com/app/transcribe). Drop in a file up to 120 minutes and 95 MiB, or record from the microphone.
2. Pick the settings: Choose Spoken or Written for numbers and Verbatim or Clean for style, and add the names the recording contains.
3. Read it beside the audio: Run it, check the names and the numbers against the recording, and copy the text or download a caption file.

[How to transcribe audio to text](https://paxalabs.com/guides/how-to-transcribe-audio-to-text) walks through a first file in detail, including how to record so the transcript comes out clean.

## What should you read or try next?

- See the product on [the speech to text page](https://paxalabs.com/speech-to-text), or start from [online transcription](https://paxalabs.com/transcribe).
- Read about live audio on [the realtime speech to text page](https://paxalabs.com/realtime-speech-to-text).
- Turn a video into captions with [the subtitle generator](https://paxalabs.com/subtitle-generator).
- Compare providers on [the speech to text API comparison](https://paxalabs.com/compare/speech-to-text-apis), or read [which API fits multilingual audio](https://paxalabs.com/guides/best-speech-to-text-api-for-multilingual-audio).
- Follow a walkthrough for [a Zoom meeting](https://paxalabs.com/guides/how-to-transcribe-a-zoom-meeting), [a YouTube video](https://paxalabs.com/guides/how-to-transcribe-a-youtube-video), [a podcast](https://paxalabs.com/guides/how-to-transcribe-a-podcast), or [an interview](https://paxalabs.com/guides/how-to-transcribe-an-interview).

## Common questions

### Is speech to text the same as ASR?

Yes. Automatic speech recognition, ASR, is the research name for the same job, and speech to text and transcription are the names products use. [The ASR glossary](https://paxalabs.com/asr) defines the terms around it.

### Is speech to text the same as voice recognition?

They are different jobs, though people often use the names loosely. Speech to text writes down what was said. Voice recognition, in its strict sense, identifies who is speaking from the sound of their voice. Paxa transcribes speech and identifies nobody.

### Does speech to text work for Thai?

Yes. Paxa STT Lite transcribes Thai, including English words spoken inside a Thai sentence, and splits Thai into words so timestamps and caption lines land on word boundaries. [The buyer's guide to Thai speech to text](https://paxalabs.com/guides/best-thai-speech-to-text) covers the other services that list Thai.

### Can speech to text tell speakers apart?

Yes, when the service offers speaker separation. Paxa STT Lite can return a recording as numbered turns, one per speaker. [The speaker turns guide](https://paxalabs.com/guides/thai-call-transcripts-with-speaker-turns) shows what that looks like on a call.

### Is speech to text free?

Phones and computers include basic dictation at no cost. For recordings, a new Paxa account's 100 free credits cover about 12 minutes of audio with no card, for personal use and evaluation.
