Does the transcript say who spoke?

It numbers the speakers. Set diarization to true and each turn arrives with a speaker number, counted from 0 by first appearance. No name is attached. Turns come from one pass over the recording. The file must be under 9 minutes, which covers a normal call. A two-channel recording can still be sent as two files.

What about calls longer than an hour?

One request takes up to 60 minutes, and under 9 with speaker turns on. Split a longer call at a silence and send the parts in parallel. Each part numbers its speakers from 0.

Can I search a call by time?

Ask for word timestamps and every word carries its start and end in seconds. A keyword match then links to the moment it was said.

What does a failed transcription cost?

Nothing. The charge is taken first and refunded automatically when the request fails, and a recording the model cannot read is refused before the charge.

All use cases

Thai speech to text for call recordings

Turn every recorded call into searchable Thai text. QA reads all of them, alerts fire on the words that matter, and a customer's history is a search away.

Try it now Read the docs

What the transcript says

A promise to pay

spokenลูกค้ายืนยันว่าจะชำระยอด สามพันสองร้อยบาท ภายในวันที่ ยี่สิบ นี้ ผ่านแอปนะคะ

writtenลูกค้ายืนยันว่าจะชำระยอด 3,200 บาท ภายในวันที่ 20 นี้ ผ่านแอปนะคะ

The written convention puts the amount and the date in digits, which is what a collections rule matches on.

A product name mid-sentence

spokenเรื่องแพ็กเกจ Fiber Max ที่สมัครไปเมื่อวานยังไม่ได้เปิดใช้เลยค่ะ

writtenเรื่องแพ็กเกจ Fiber Max ที่สมัครไปเมื่อวานยังไม่ได้เปิดใช้เลยค่ะ

The English name stays English under both conventions. A vocabulary keeps its spelling consistent across thousands of calls.

A refund request

spokenขอคืนเงินค่าบริการเดือนสิงหาคมจำนวน ห้าร้อยเก้าสิบเก้าบาท ได้ไหมคะ

writtenขอคืนเงินค่าบริการเดือนสิงหาคมจำนวน 599 บาท ได้ไหมคะ

A keyword rule on the word for refund reads the same text a person would, with the amount ready to post.

Every transcript pair on this page is an authored illustration of what the convention option changes, re-checked against live output before publish. The Speech-to-Text page carries the model's own witnessed transcripts of a real recording, with the capture date.

How it fits a contact centre

A contact centre records every call and listens to almost none of them. A QA team samples a handful a week, scores them by ear, and the rest sit in storage until a complaint sends someone digging. Transcription changes the arithmetic. Every call becomes text the moment it ends. A script check reads all of them. A keyword alert fires on the calls that mention a refund, a competitor, or a promise to pay. When a customer calls back, the agent reads the last conversation and skips asking for it again. The transcript keeps what was said in the language it was said in. An English product name inside a Thai sentence survives the trip.

Send each recording when the call endsPost the file the telephony system already writes, up to 60 minutes. The length is read before the charge. A corrupt recording costs nothing.
Ask for the written conventionAmounts, dates, and times arrive as digits. The rules that match on them stay simple. Keep the spoken convention for a verbatim record.
Run your rules over the textScript checks, keyword alerts, and customer search are yours. A vocabulary of up to 50 product and plan names keeps the spelling consistent.

What every recorded call becomes

100%
Script coverage: A greeting, an identity check, and a closing line are string matches on the transcript. Every call is checked, and the ones that miss a line are the ones QA listens to.; 0 and 1
Speakers: With diarization on, each turn carries a speaker number, counted from 0 by first appearance. A script check finds the agent's turns by their number and their wording.
Keyword alerts: Refund, cancel, complaint, a competitor's name: a rule over the transcript raises the call within minutes of it ending, while the customer still remembers it.
Customer history: Index the transcripts by customer. The next agent reads the last call before answering, and the customer tells the story once.

The model writes what was said. Scoring, alerts, and routing are rules you write over the text.

A script check, as a wire log

qa-check.log

call  2026-09-08T10:14:02+07:00  agent=A17  duration=4:12

greeting     ✓  0.4s   สวัสดีค่ะ ... ยินดีให้บริการค่ะ
identity     ✓  11.8s  ขอทราบเลขบัตรประชาชน 4 ตัวท้าย
hold-notice  ✗         (no match)
closing      ✓  241.9s ขอบคุณที่ใช้บริการค่ะ

keywords     คืนเงิน @ 96.2s   ยกเลิก @ 133.0s
flag         review: missing hold notice, refund mentioned

Word timestamps give each match its second. A reviewer jumps straight to it. The times shown are illustrative.

What a month of calls costs

Workload	Audio	Credits
One call, 4 minutes	4 min	33.34
A day, 100 calls	7 h	3,334
A month, 2,000 calls		66,680 credits

8.33 credits a minute of audio, billed by the millisecond. The transcript options change nothing.

For developers

The snippets send one recording with the written convention and read the transcript and the credits charged from the response.

AUDIO=$(base64 < meeting.m4a | tr -d '\n')
curl -X POST https://api.paxalabs.com/v1/stt \
  --max-time 300 \
  -H "Authorization: Bearer $PAXA_API_KEY" \
  -H "Content-Type: application/json" \
  -d "{\"audio\": \"$AUDIO\", \"model\": \"paxa-stt-lite-v1-preview\", \"convention\": \"written\"}"

import { readFile } from "node:fs/promises";

const audio = (await readFile("meeting.m4a")).toString("base64");

const response = await fetch("https://api.paxalabs.com/v1/stt", {
  method: "POST",
  headers: {
    Authorization: `Bearer ${process.env.PAXA_API_KEY}`,
    "Content-Type": "application/json",
  },
  body: JSON.stringify({ audio, model: "paxa-stt-lite-v1-preview" }),
  // A long recording can run for minutes; give it room.
  signal: AbortSignal.timeout(300_000),
});

if (!response.ok) throw new Error(`Transcription failed: ${response.status}`);
const { text, usage } = await response.json();
console.log(text, `${usage.credits} credits`);

import base64
import os

import requests

with open("meeting.m4a", "rb") as file:
    audio = base64.b64encode(file.read()).decode()

response = requests.post(
    "https://api.paxalabs.com/v1/stt",
    headers={"Authorization": f"Bearer {os.environ['PAXA_API_KEY']}"},
    json={"audio": audio, "model": "paxa-stt-lite-v1-preview"},
    # A long recording can run for minutes; give it room.
    timeout=300,
)
response.raise_for_status()
body = response.json()
print(body["text"], body["usage"]["credits"], "credits")

Full API reference

What the model was built to hear

Paxa STT Lite transcribes Thai and English, including sentences that switch language halfway. It labels speaker turns by number on request. It publishes no accuracy figure until the evaluation method is published beside it.

Common questions

Does the transcript say who spoke?: It numbers the speakers. Set diarization to true and each turn arrives with a speaker number, counted from 0 by first appearance. No name is attached. Turns come from one pass over the recording. The file must be under 9 minutes, which covers a normal call. A two-channel recording can still be sent as two files.
What about calls longer than an hour?: One request takes up to 60 minutes, and under 9 with speaker turns on. Split a longer call at a silence and send the parts in parallel. Each part numbers its speakers from 0.
Can I search a call by time?: Ask for word timestamps and every word carries its start and end in seconds. A keyword match then links to the moment it was said.
What does a failed transcription cost?: Nothing. The charge is taken first and refunded automatically when the request fails, and a recording the model cannot read is refused before the charge.

Start with free credits

Start free

More use cases