# Text to speech that speaks English, Thai, and Mandarin

Paxa TTS Flash v2 speaks English, Thai, and Mandarin in all 50 voices, even in the same sentence. You write the line and how it should sound, at 10 credits per 1,000 characters, the same price as v1.

Text-to-Speech · Flagship

- Languages: EN · TH · ZH
- Voices: 50
- Per 1M characters: $10

[Read the docs](https://paxalabs.com/docs/text-to-speech.md) · [API reference](https://paxalabs.com/docs/api/text-to-speech.md)

## Built in Bangkok for English, Thai, and Mandarin

Every Paxa TTS Flash v2 clip on this page is the model's unedited output.

## Hear text to speech doing real work

Move from six distinct voices to mixed-language sentences, contextual numbers, and complete service announcements.

- Khanom Krok, the elder: “สมัยผมยังหนุ่ม แถวนี้ยังเป็นทุ่งนา เดินไปตลาดทีหนึ่งก็ค่อนวันแล้วครับ” https://paxalabs.com/audio/paxa-tts-flash-v2/khanomkrok.m4a
- Nom Yen, with word timing: “ยินดีต้อนรับสู่ Paxa Labs ค่ะ เสียงภาษาไทยที่สลับเป็น English ได้กลางประโยค อย่างเป็นธรรมชาติ” https://paxalabs.com/audio/paxa-tts-flash-v2/nomyen.m4a
- Massaman, the leading man: “ในเมืองที่ไม่เคยหลับ มีเรื่องราวหนึ่งรอให้ใครสักคนกล้าเล่ามันออกมา” https://paxalabs.com/audio/paxa-tts-flash-v2/massaman.m4a
- Foi Thong, the narrator: “เรื่องราวทุกเรื่องมีจุดเริ่มต้น และวันนี้ ฉันจะเล่าให้คุณฟังตั้งแต่ต้นจนจบค่ะ” https://paxalabs.com/audio/paxa-tts-flash-v2/foithong.m4a
- Latte, the brand film voice: “Every great brand has a story. Tonight, we're going to tell yours.” https://paxalabs.com/audio/paxa-tts-flash-v2/latte.m4a
- Tao Huay, the drama lead: “如果时间可以倒流，我还是会在那个下雨的傍晚，安静地等你回来。” https://paxalabs.com/audio/paxa-tts-flash-v2/taohuay.m4a
- A tracking number, mid-sentence: “ออเดอร์ของคุณกำลังจัดส่งโดย Flash Express หมายเลขพัสดุ TH123456789 ค่ะ” https://paxalabs.com/audio/ideas/voice-agents-3.m4a
- A vocabulary lesson, both directions: “คำศัพท์วันนี้คือ negotiate แปลว่า เจรจาต่อรอง ลองฟังตัวอย่างประโยคนะคะ We need to negotiate the contract terms.” https://paxalabs.com/audio/ideas/e-learning-3.m4a
- An IVR line that switches for one sentence: “ขออภัยค่ะ เจ้าหน้าที่ทุกท่านกำลังให้บริการ For English, please press 9. กรุณาถือสายรอสักครู่นะคะ” https://paxalabs.com/audio/ideas/call-centers-4.m4a
- An amount and a date: “สวัสดีค่ะ คุณสมชาย ยอดชำระของท่านจำนวน 1,250 บาท ครบกำหนดวันที่ 15 สิงหาคมนี้ ชำระผ่านแอปได้ตลอด 24 ชั่วโมงค่ะ” https://paxalabs.com/audio/ideas/call-centers-2.m4a
- A four-digit balance: “ยอดคงเหลือในบัญชีของคุณคือ 8,432 บาทค่ะ ต้องการทำรายการอื่นเพิ่มไหมคะ” https://paxalabs.com/audio/ideas/voice-agents-2.m4a
- A queue call: “หมายเลข 47 เชิญช่องบริการที่ 3 ค่ะ” https://paxalabs.com/audio/ideas/announcements-2.m4a
- A bilingual safety announcement: “เพื่อความปลอดภัย กรุณายืนหลังเส้นเหลือง For your safety, please stand behind the yellow line.” https://paxalabs.com/audio/ideas/announcements-4.m4a
- A news lede, phrased without punctuation: “ข่าวด่วน กรมอุตุนิยมวิทยาเตือนฝนตกหนักในพื้นที่กรุงเทพฯ และปริมณฑลช่วงเย็นวันนี้” https://paxalabs.com/audio/ideas/publishers-3.m4a
- The full service call: “สวัสดีค่ะ ขอบคุณที่ติดต่อบริษัทของเรา กด 1 เพื่อติดต่อฝ่ายขาย กด 2 ฝ่ายบริการลูกค้า หรือกด 0 เพื่อติดต่อเจ้าหน้าที่ค่ะ” https://paxalabs.com/audio/ideas/call-centers-1.m4a

## Write the emotion into the text

Every Paxa TTS Flash v2 voice reads direction tags written into the text, such as [happy], [calm], or [sad:0.6].

- Tagged lines, as a product writes them: “[happy] สอบผ่านแล้ว [sad:0.6] แต่คงคิดถึงที่นี่” https://paxalabs.com/audio/emotion-scenes/request.m4a. Each line is the text of one request, sent as printed with its tags. The clip is the audio that request returned.
- A companion's reply, three tags in one turn: “[happy] ได้ข่าวว่าสอบผ่านแล้ว ดีใจด้วยจริง ๆ นะ [sad:0.6] แต่ก็ใจหายเหมือนกัน ต่อไปคงไม่ค่อยได้เจอกันแล้ว [calm] ไม่เป็นไร เดี๋ยวเรานัดกินข้าวกันนะ” https://paxalabs.com/audio/emotion-scenes/companion.m4a. I heard you passed, I'm really happy for you. It's sad too, we won't see each other much now. It's fine, let's get dinner soon.

## The reading rules of Thai, written down

Thai spelling does not map cleanly from character to sound. Each row shows the expected reading and the context a useful voice model has to recover.

### Syllables the spelling hides

- กด (กด, kot): The vowel is real but never written; nothing in the string marks it.
- สรร · กรรม (สัน · กัม, san · kam): The doubled -รร reads un at the end of a word but um before a final; the reader must look ahead.
- ขรม (ขะ-หฺรม, kha-rom): Officially two syllables: a leading consonant, not a true cluster like ปลา.
- ปลา · ปลัด (ปลา · ปะ-หลัด, plaa · pa-lat): The same two letters begin a cluster in one word and two syllables in the other.
- รัฐศาสตร์ (รัด-ถะ-สาด, rat-tha-saat): A linking syllable appears with no letter of its own to trigger it.
- จักรยาน · จักรี (จัก-กระ-ยาน · จัก-กรี, jak-kra-yaan · jak-kri): The same letters cut into different syllables; only the word decides.
- แพทย์ · จันทร์ (แพด · จัน, phaet · jan): The cancellation mark deletes letters; จันทร์ drops two at once.

### Marks, repeats, and numerals

- เด็กเล็ก ๆ (เด็ก-เล็ก-เล็ก, dek-lek-lek): ๆ repeats one word here, two syllables in ในวันหนึ่ง ๆ, and a whole phrase in แต่ละวัน ๆ; meaning sets the span.
- กรุงเทพฯ (กรุง-เทบ-มะ-หา-นะ-คอน, krung-thep-maha-nakhon): ฯ restores six syllables here; royal formulas repeat the visible part instead (โปรดเกล้าฯ).
- ฯลฯ (และอื่น ๆ · ละ-ถึง, lae uen uen · la-thueng): At the end of a list it reads one way; between two items (ก ฯลฯ ฮ) it reads through-to.
- ก.พ. (กุม-พา-พัน หรือ กอ-พอ, kumphaphan or ko-pho): Two legal readings sit side by side. One reads the full word. The other reads the letter names, the form accepted for long abbreviations.
- ๑๐๑ · 10150 (ร้อย-เอ็ด · หนึ่ง-ศูนย์-หนึ่ง-ห้า-ศูนย์, roi-et · digit by digit): A quantity takes เอ็ด for a trailing one; a postal code reads digit by digit; ๑๖.๓๐ น. reads as a clock time.
- ๒๕๖๙ · 2026: Thai and Arabic digits live in the same sentence; the Thai block carries its own zero through nine.

Real text breaks these rules constantly. A useful reader has to handle both the prescribed form and what people actually type.

- **Phrasing comes from context**: Thai usually arrives without spaces between words. Flash finds phrase boundaries from the sentence. Long passages breathe naturally, with every pause on a phrase boundary and none at a fixed character interval.
- **Pronunciation is contextual**: The same letters can split into different syllables, marks can suppress sounds, and numerals change reading by use. Flash resolves the complete word and sentence before it speaks.
- **English stays in character**: Product names, tracking codes, and full English clauses stay in the same voice as the surrounding Thai, keeping one speaker and one delivery throughout.

## Timed speech

Nom Yen reading the cast line, with the word spans the timestamps option returns.

## Build with the text to speech API

- **HTTP streaming**: Set stream to true and chunked audio starts arriving before synthesis finishes.
- **OpenAI-compatible**: Point the OpenAI SDK at /v1/audio/speech and keep the rest of your code.
- **Word timestamps**: Return word, sentence, or utterance spans for captions and read-along interfaces.

## The hard parts of Thai, read aloud

**Code-switching**: “เมื่อเช้ามี meeting กับ team ตั้งแต่เก้าโมง ตอนแรกกะจะ present slide ที่เตรียมไว้ แต่ boss ขอ move ไปคุยเรื่อง budget ก่อน เพราะ client เพิ่งส่ง feedback มาว่าอยากได้ timeline ที่ clear กว่านี้ สรุปคือทั้ง team ต้อง revise plan ใหม่หมด แล้ว submit ภายในวันศุกร์ เครียดมาก ขนาด lunch ยังไม่ได้กินเลย บ่ายนี้ก็มี call กับ vendor อีกรอบ สงสัยต้องสั่ง coffee มา boost พลังก่อนแล้วล่ะ” Real Thai office speech switches to English mid-sentence. Here meeting, budget, feedback, and timeline all sit inside Thai grammar, and the reader has to carry both languages in one voice without breaking stride.

- Paxa (paxa-tts-flash-v1): https://paxalabs.com/audio/compare/codeswitch-paxa.m4a
- ElevenLabs (Eleven v3): https://paxalabs.com/audio/compare/codeswitch-elevenlabs.m4a
- Google Gemini (Gemini 3.1 Flash TTS): https://paxalabs.com/audio/compare/codeswitch-gemini.m4a
- Cartesia (Sonic 3.5): https://paxalabs.com/audio/compare/codeswitch-cartesia.m4a
- MiniMax (speech-2.8-hd): https://paxalabs.com/audio/compare/codeswitch-minimax.m4a
- iApp (ChindaTTS): https://paxalabs.com/audio/compare/codeswitch-iapp.m4a
- Narakeet (default voice): https://paxalabs.com/audio/compare/codeswitch-narakeet.m4a
- Botnoi Voice (default voice): https://paxalabs.com/audio/compare/codeswitch-botnoi.m4a

Every provider read the same text. The clips are unedited apart from volume matching for fair listening and transcoding to AAC for delivery, and the model behind each clip is named on its row.

## What text to speech costs

- `paxa-tts-flash-v1` (Paxa TTS Flash): 10 credits per 1,000 characters, `max_chars` 5,000, `max_stream_chars` 20,000.
- `paxa-tts-flash-v2` (Paxa TTS Flash v2): 10 credits per 1,000 characters, `max_chars` 5,000, `max_stream_chars` 20,000.

Costs shown at cent precision. Credits are shared across every Paxa API.

Start with 100 free credits, no card required. When you outgrow them, Starter is $6 a month for 8,100 credits, and refill packs never expire.

## Questions, answered

### How natural do the voices sound?

Listen to the clips, then paste your own text in the playground and judge with your ears. New accounts get 100 free credits to test the voices on material you know.

### Can it switch between Thai and English mid-sentence?

Yes. One request can carry one voice across both languages, including product names, tracking numbers, and complete English clauses. The clips above include three examples.

### What does an hour of speech cost?

About $0.54 in English and $0.36 in Thai. The rate is 10 credits per 1,000 characters, and an hour of spoken English runs about 53,820 characters against about 36,360 for Thai. Failed requests refund automatically.

### Can I try it for free?

Yes. New accounts get 100 credits with no card, which converts about 10,000 characters, roughly 11 minutes of spoken English or 16 of Thai. After that the rate is 10 credits per 1,000 characters.

### How does it read Thai numbers and dates?

It uses the surrounding context. Quantities, identifiers, postal codes, and clock times each receive their own reading. The clips above let you check amounts, dates, balances, and a queue number yourself.

### What makes an AI TTS model good at Thai?

Thai speech synthesis has to recover syllable boundaries, tone, and the reading a word takes in context before it can produce a sound. Flash reads Thai as a first language and resolves all three from the sentence around them. The reading rules above collect the cases that catch a per-character reader.

### Can I stream the audio?

Yes, in three ways. Chunked HTTP streaming gives fast first audio, an OpenAI-compatible endpoint works with the SDK you already use, and a live websocket speaks while your LLM is still writing.

### Does it work with the OpenAI SDK?

Yes. Point the SDK's base URL at the API and call /v1/audio/speech; requests stream by default and bill at the same per-character rate.

### Can I get word-level timestamps for captions?

Yes. Ask for word, sentence, or utterance spans. The Timed speech section renders the same shape of data as a live read-along highlight.

### Can I change the speaking speed?

Yes. The speed field takes a multiplier from 0.5 to 1.5 on the pace a voice was designed for. It has no effect on what a request costs, because the price follows the character count.

### Can the voice sound happy or sad?

Yes. Every Paxa TTS Flash v2 voice reads direction tags written into the text, such as [happy] or [sad:0.6]. Tag characters bill like the rest of the text. The emotion section above plays two tagged lines.

### Can I use the audio commercially?

Yes, on any paid plan or credit purchase. The Terms assign the generated audio to you, including audio you generated on free credits before your first purchase. Use it in your product, your videos, or your phone system. Free credits on their own cover personal projects and evaluation.

## Changelog

- 2026-11-17, Many live sessions on one WebSocket: WS /v1/tts/multi carries many live speech sessions on one WebSocket, for an application that speaks on many calls at once. Each session is a context you name with a context_id. It takes the frames of /v1/tts/live, with the same billing and limits. Audio frames open with a header naming their context. A context ends with a closed event unless the socket closes first, and the others keep running. Each open context holds one of the plan's concurrent speech slots.
- 2026-11-04, Longer streamed requests: A streamed request to Paxa TTS Flash or Paxa TTS Flash v2 on POST /v1/tts or POST /v1/audio/speech now takes up to 20,000 characters, reported as max_stream_chars on GET /v1/models. A long article or script goes out in fewer requests, and the console's Speech tool streams the same length. A buffered request and each text frame of the live socket stay at 5,000. The rate is unchanged.
- 2026-11-03, Multi-speaker dialogue: POST /v1/tts/dialogue speaks up to 100 lines, each in its own voice, as one streamed audio file, with gap_ms of silence between lines, 300 ms by default. The lines hold up to 20,000 characters in all and bill as one request. A refusal before the charge names the line at fault, and a failure after it keeps only the price of the lines delivered whole. Scripts in the console's Speech tool now run as one dialogue.
- 2026-10-31, Streamed speech starts sooner: Streamed speech on POST /v1/tts, POST /v1/audio/speech, and the live socket now starts its audio sooner. A stream can now be refused after its audio began. It then ends early, and nothing is charged. An NDJSON stream ends with an error event, and a raw stream closes before its end. A live leg sends error and closes the connection, and its charged event can now follow its first audio.
- 2026-10-28, Temperature and pause tags: Paxa TTS Flash v2 reads temperature on POST /v1/tts, POST /v1/audio/speech, and the live start frame. It runs from 0 to 2, default 0.5, and sets how much a reading varies from one request to the next. A request that leaves it out reads at 0.5, which varies less than before. At 0 the same text in the same voice varies least. The console's Speech tool sets it with a slider. [short pause], [pause], and [long pause] add silence after the word before them, and [pause 300ms] sets an exact length up to 500 ms. A longer one plays at 500 ms. [emphasis] now reads as a direction.
- 2026-10-26, sample_rate and picker lists for OpenAI-compatible apps: POST /v1/audio/speech reads sample_rate for pcm and wav as POST /v1/tts does, and ignores it with other formats, though any value must be one of the listed rates. A voice platform that plays raw pcm at a fixed rate gets audio at that rate, and on Paxa TTS Flash, which serves only 24000 Hz, another rate answers unsupported_format. GET /v1/audio/voices and GET /v1/audio/models list Paxa TTS Flash v2's voices and the model itself, with no key, at the paths OpenAI-compatible apps such as Open WebUI read for their pickers. The live socket's started event now names the format and sample_rate its audio arrives in.
- 2026-10-24, OpenAI's names, raw pcm, and phone-line audio: POST /v1/audio/speech accepts OpenAI's model names, tts-1, tts-1-hd, gpt-4o-mini-tts, gpt-4o-mini-tts-2025-03-20, gpt-4o-mini-tts-2025-12-15, which run on paxa-tts-flash-v2, and its 13 voice names, each mapped to a Paxa TTS Flash v2 voice. The x-voice-id response header names the voice that spoke. Every speech surface serves raw pcm, and Paxa TTS Flash v2 serves pcm and wav at 8 to 48 kHz through sample_rate, plus G.711 mulaw and alaw at 8 kHz for phone lines.
- 2026-10-23, Design voices through the API: POST /v1/voices/design makes a voice from a description, or refines a draft with an instruction, and returns the voice's words and its line as MP3 for 5 credits a call. POST /v1/voices/designed saves a draft, GET /v1/voices/designed lists the account's voices and slots, and PATCH and DELETE /v1/voices/designed/{id} rename and delete one. Drafts from the console and the API save in either place.
- 2026-10-21, Design your own voice: The Voice design tool in the console turns a description in words into a voice and speaks a line in it. Ask for a change, and it answers with the changed voice. Each design or refine turn costs 5 credits. A saved voice gets an id starting voice_, which works on POST /v1/tts, the OpenAI-compatible route, and live streaming with Paxa TTS Flash v2, for your account only. Speech in it costs 1.25 times the standard rate, or 1.5 times with timestamps. Free keeps 1 voice, Starter 3, Standard 10, and Pro 50.
- 2026-10-16, Paxa TTS Flash v2 is available now: Paxa TTS Flash v2 (paxa-tts-flash-v2) is available on POST /v1/tts, the OpenAI-compatible route, live streaming, and the Speech tool in the console. A direction in square brackets, written in English, Thai, or Chinese, sets how the words after it are delivered, and sounds, hesitations, and <emph> and <whisper> marks play where they are written. Its 50 voices each speak English, Thai, and Mandarin, at the 10 credits per 1,000 characters of Paxa TTS Flash.
- 2026-10-01, Speech costs a third less: Paxa TTS Flash now costs 10 credits per 1,000 characters, which is $10 per million, down from 15 credits. The new rate applies to every speech request and to the Speech tool in the console, and timestamps and live streaming fall in the same proportion. Credits already in a balance or a plan now voice 50 percent more text.
- 2026-10-01, A language setting for speech: A new language setting lets the voice read Thai, English, and Mandarin Chinese text more accurately. Set it to the language your text is written in, and numbers, symbols, and codes such as AB2039 are read to suit that language. Auto, the default, works it out from the text. The setting works on POST /v1/tts, live streaming, and the Speech tool in the console, and does not change the charge.
- 2026-09-25, Emotion tags on two experimental voices: Khanom Chan and Kaprao join Paxa TTS Flash as experimental voices that read emotion tags written into the text, bringing the roster to 34. There are 26 tags, such as [happy], [sad], and [furious], each with an optional weight from 0 to 1, as in [sad:0.5]. Tag characters bill as text, every other voice skips the tags, and GET /v1/voices adds experimental and emotion_tags.
- 2026-09-14, Mandarin Chinese and six new voices: Paxa TTS Flash now reads Mandarin Chinese in Simplified script beside Thai and English. The roster grows to 32 voices with Tao Huay and Oolong in Mandarin, Espresso and Mocha in English, and Khao Niao and Tom Yum in Thai. The language field of GET /v1/voices adds "zh".
- 2026-09-06, MCP server for agents: @paxalabs/mcp gives Claude Code, Claude Desktop, Cursor, and other MCP clients the API as tools: speech through the machine's speakers, speech to file, translation into Thai, and OCR. A paxa say hook speaks up when Claude Code needs you.
- 2026-09-01, Interrupt a live reply: A cancel frame on /v1/tts/live stops the leg being spoken and drops whatever text is still buffered. An assistant can stop the moment the speaker talks over it. The stopped leg keeps the charge its receipt reported; the discarded buffer was never charged.
- 2026-08-09, Timestamps and live streaming: Requests can return word, sentence, or utterance timing, and the /v1/tts/live websocket streams speech while your text is still arriving.
- 2026-08-03, The speech API ships: Paxa TTS Flash (paxa-tts-flash-v1) opens to the public through POST /v1/tts: Thai and English speech with charge-first billing and automatic refunds on failed requests.

## Start free: 100 credits of text to speech

No card and no sales call. 100 credits convert about 10,000 characters, roughly 11 minutes of spoken English in any of the 50 voices. Sign in, paste your text, and hear it within a minute.

## Related reading

- [What a Thai voiceover costs in 2026: studio, freelancer, AI](https://paxalabs.com/guides/thai-voice-over-cost.md)
- [Choosing a Thai AI voice for the job, across all 50 voices](https://paxalabs.com/guides/choosing-a-thai-ai-voice.md)
- [How to write a script for a Thai AI voice](https://paxalabs.com/guides/writing-for-thai-text-to-speech.md)
- [Thai has no full stop: where streaming speech should break](https://paxalabs.com/guides/thai-sentence-breaks-for-streaming-speech.md)
- [How to judge a Thai AI voice in ten sentences](https://paxalabs.com/guides/judging-a-thai-ai-voice.md)
- [Paxa TTS Flash now speaks Mandarin Chinese, with six new voices](https://paxalabs.com/blog/mandarin-chinese-and-six-new-voices.md)
- [Paxa TTS Flash now reads emotion tags, on two new voices](https://paxalabs.com/blog/emotion-tags-for-thai-speech.md)
- [Paxa TTS Flash starts speaking in 104 milliseconds](https://paxalabs.com/blog/tts-latency-singapore-bangkok.md)
- [What is text to speech? How AI voices work and where they are used](https://paxalabs.com/guides/what-is-text-to-speech.md)
- [How to add an AI voiceover to TikTok, Reels, and Shorts](https://paxalabs.com/guides/ai-voiceover-for-tiktok-reels-and-shorts.md)
- [Best Thai text to speech for apps, videos, and phone lines](https://paxalabs.com/guides/best-thai-text-to-speech.md)
- [Best AI voice generator for YouTube narration, faceless channels, and Shorts](https://paxalabs.com/guides/best-ai-voice-generator-for-youtube.md)
- [Best free text to speech for personal projects and trying voices before you pay](https://paxalabs.com/guides/best-free-text-to-speech.md)
- [Best text to speech for audiobooks and long-form narration](https://paxalabs.com/guides/best-text-to-speech-for-audiobooks.md)
- [How to make an AI voice, from script to download](https://paxalabs.com/guides/how-to-make-an-ai-voice.md)
- [How to have AI read a novel aloud](https://paxalabs.com/guides/ai-reads-novels-aloud.md)
- [How to describe a voice for AI voice design, with examples](https://paxalabs.com/guides/how-to-describe-a-voice-for-ai.md)
- [Voice design vs voice cloning for a voice of your own](https://paxalabs.com/guides/voice-design-vs-voice-cloning.md)
- [Best AI voice design tools for Thai voices, saved voices, and voice agents](https://paxalabs.com/guides/best-ai-voice-design-tools.md)
- [What is the difference between text to speech and speech to text?](https://paxalabs.com/guides/text-to-speech-vs-speech-to-text.md)
- [What is an AI voice agent? How one works and where it is used](https://paxalabs.com/guides/what-is-an-ai-voice-agent.md)
- [Best text to speech APIs for voice agents](https://paxalabs.com/guides/best-text-to-speech-api-for-voice-agents.md)
- [Best Chinese text to speech for apps, videos, and voice agents](https://paxalabs.com/guides/best-chinese-text-to-speech.md)
- [How to dub a video into another language with AI](https://paxalabs.com/guides/how-to-dub-a-video-with-ai.md)
- [Paxa TTS Flash v2 is available now](https://paxalabs.com/blog/paxa-tts-flash-v2.md)
- [Voice design is available now](https://paxalabs.com/blog/voice-design.md)
- [Temperature and pause tags on Paxa TTS Flash v2](https://paxalabs.com/blog/temperature-and-pause-tags.md)
- [Speech and transcription for many live calls on one WebSocket](https://paxalabs.com/blog/many-calls-on-one-websocket.md)

## Research behind it

- [A Thai voice from fifteen seconds of speech](https://paxalabs.com/research/thai-tts-from-synthetic-speech.md)
- [CPU speech synthesis under a serverless bill](https://paxalabs.com/research/cpu-tts-under-serverless-billing.md)

## More from the lab

- [Speech-to-Text](https://paxalabs.com/speech-to-text.md)
- [OCR and Extraction](https://paxalabs.com/ocr.md)
- [AI Translation](https://paxalabs.com/translation.md)
