xAI TTS v1 API by xAI
xAI TTS v1 is a high-fidelity text-to-speech model that converts text into natural, expressive speech with sub-second latency, supporting 20 languages and 80+ voices with fine-grained delivery control.
xAI TTS v1 do xAI phát triển. Atlas Cloud (vận hành bởi Atlas Cloud AI LLC) cung cấp quyền truy cập vào mô hình này và không sở hữu mô hình. Mọi thương hiệu đều thuộc về chủ sở hữu tương ứng.
xAI TTS v1 — Text to Speech
Developer: xAI
Model ID: xai/tts-v1
Release Date: April 2026
Overview
xAI TTS v1 is a high-fidelity text-to-speech model developed by xAI, the company behind Grok. It converts text into natural, expressive speech with sub-second latency, supporting 20 languages and a rich voice library of 80+ voices. The model features fine-grained delivery control through inline speech tags, multiple audio output formats for both streaming and telephony use cases, and a voice cloning capability that can produce a production-ready custom voice from roughly a minute of reference audio.
xAI TTS v1 is designed for real-world production workloads — customer service bots, content narration, accessibility tools, and real-time voice agents — offering competitive per-character pricing with enterprise-grade compliance.
Key Capabilities
- Expressive delivery control — 14 instant speech tags and 13 wrapping style tags for pauses, laughter, whispers, pitch shifts, speed changes, and more.
- 80+ voices across 20 languages — Five universal multilingual voices plus language-optimized voices for Chinese, Russian, Italian, French, Spanish, Hindi, Japanese, Korean, Portuguese, German, Dutch, Polish, Turkish, Arabic, Vietnamese, Thai, Danish, Swedish, Finnish, and English.
- Flexible audio output — MP3, WAV, PCM, μ-law, and A-law codecs with sample rates from 8 kHz (telephony) to 48 kHz (studio).
- Voice cloning — Record up to 120 seconds of reference audio; a custom
voice_idis ready in under 2 minutes at no additional cost. - Streaming and batch modes — Both a standard REST endpoint and a WebSocket streaming endpoint for real-time audio delivery.
- Automatic language detection — Set
language: "auto"to let the model identify the input language. - Privacy-first — Audio is never stored or used for model training.
Use Cases
- Customer service / IVR — Low-latency telephony output (μ-law/A-law at 8 kHz) for automated voice response systems.
- Content narration — Podcast episodes, audiobooks, and video voiceovers with expressive, human-like delivery.
- Accessibility — Real-time screen readers and reading assistants for visually impaired users.
- Real-time voice agents — Paired with xAI's Speech-to-Text and Voice Agent APIs for end-to-end conversational AI.
- Multilingual applications — Localized content delivery across 20 languages from a single API.
- Custom brand voices — Clone a spokesperson's voice to maintain consistent audio identity at scale.
Voices
Multilingual Voices
These five voices support all 20 available languages.
| Voice ID | Name | Gender | Character |
|---|---|---|---|
eve | Eve | Female | Energetic, upbeat — default voice |
ara | Ara | Female | Warm and conversational |
leo | Leo | Male | Authoritative, instructional |
rex | Rex | Male | Professional, business tone |
sal | Sal | Male | Versatile, neutral |
Language-Optimized Voices (selected)
Language-specific voices are optimized for their native language and automatically lock the language field when selected.
| Language | Voices |
|---|---|
| English | Grace (F), Claire (F), James (M), Daniel (M) |
| Chinese (Mandarin) | Jian (M), Hao (F), Xia (F) |
| Russian | Pavel (M), Andrei (M), Dmitri (M), Irina (F) |
| French | Remi (M), Hugo (M), Camille (F) |
| Spanish | Manuel (M), Javier (M), Diego (M), Andres (M) |
| Italian | Enzo (M), Matteo (M), Alessandro (M), Luca (F) |
| German | Moritz (M), Niklas (M), Clara (F), Lena (F) |
| Japanese | Ren (M), Sakura (F) |
| Korean | Jun-seo (M), Min-jun (M), Seo-yeon (F), Ji-yeon (F) |
| Portuguese | Mateus (M), Rafael (M), Beatriz (F) |
| Hindi | Karan (M), Ananya (F) |
| Arabic | Khalid (M), Tariq (M), Layla (F) |
| Turkish | Emre (M), Kaan (M), Aylin (F) |
| Dutch | Thijs (M), Ruben (M), Femke (F), Noor (F) |
| Polish | Mateusz (M), Jakub (M), Katarzyna (F), Aleksandra (F) |
| Vietnamese | Duc (M), Minh (M), Mai (F) |
| Swedish | Axel (M), Erik (M), Saga (F) |
| Finnish | Valtteri (M), Eero (M), Helmi (F), Elina (F) |
| Thai | Krit (M), Aroon (M) |
| Danish | Lars (M), Kasper (M), Ida (F) |
The complete voice library, including custom cloned voices, is accessible via the
voice_idparameter.
Supported Languages
| Code | Language |
|---|---|
auto | Auto Detect |
en | English |
zh | Chinese (Mandarin) |
ar-EG | Arabic (Egypt) |
ar-SA | Arabic (Saudi Arabia) |
ar-AE | Arabic (UAE) |
bn | Bengali |
fr | French |
de | German |
hi | Hindi |
id | Indonesian |
it | Italian |
ja | Japanese |
ko | Korean |
pt-BR | Portuguese (Brazil) |
pt-PT | Portuguese (Portugal) |
ru | Russian |
es-MX | Spanish (Mexico) |
es-ES | Spanish (Spain) |
tr | Turkish |
vi | Vietnamese |
Input Parameters
| Parameter | Type | Default | Description |
|---|---|---|---|
model | string | xai/tts-v1 | Model identifier. Required. |
text | string | — | Text to synthesize. Maximum 15,000 characters. Supports inline speech tags. Required. |
language | string | auto | BCP-47 language code or "auto" for automatic detection. Required. |
voice_id | string | eve | Voice identifier. Multilingual voices support all languages; language-optimized voices lock to their native language. |
codec | string | mp3 | Output audio codec: mp3, wav, pcm, mulaw, alaw. |
sample_rate | integer | 24000 | Sample rate in Hz. Options: 8000, 16000, 22050, 24000, 44100, 48000. |
bit_rate | integer | 128000 | Bit rate in bps. MP3 only. Options: 32000, 64000, 96000, 128000, 192000. |
speed | number | 1.0 | Playback speed multiplier. Range: 0.7 (slower) to 1.5 (faster). |
text_normalization | boolean | false | When true, converts written-form text (numbers, abbreviations, symbols) to spoken form before synthesis. |
optimize_streaming_latency | integer | 0 | Streaming latency trade-off: 0 = best quality, 1 = lower first-chunk latency, 2 = lowest first-chunk latency. |
Audio Format Recommendations
| Use Case | Codec | Sample Rate | Bit Rate |
|---|---|---|---|
| Standard playback (default) | mp3 | 24,000 Hz | 128,000 bps |
| Studio / high-fidelity | mp3 or wav | 44,100 or 48,000 Hz | 192,000 bps |
| Telephony / IVR | mulaw or alaw | 8,000 Hz | — |
| Streaming (web) | mp3 | 22,050 Hz | 64,000 bps |
Inline Speech Tags
Speech tags let you control delivery at a granular level within the text field.
Instant Tags
Insert at any point in the text to trigger an immediate vocal event.
| Tag | Effect |
|---|---|
[pause] | Brief pause |
[long-pause] | Extended pause |
[hum-tune] | Soft humming sound |
[laugh] | Full laugh |
[chuckle] | Short chuckle |
[giggle] | Light giggle |
[cry] | Crying sound |
[tsk] | Disapproving tsk |
[tongue-click] | Tongue click |
[lip-smack] | Lip-smack sound |
[breath] | Audible breath |
[inhale] | Inhalation |
[exhale] | Exhalation |
[sigh] | Sigh |
Wrapping Style Tags
Wrap text to apply a delivery style to the enclosed span.
| Tag | Effect |
|---|---|
<soft>…</soft> | Softer, quieter delivery |
<whisper>…</whisper> | Whispered speech |
<loud>…</loud> | Louder, more projected delivery |
<build-intensity>…</build-intensity> | Gradually increasing intensity |
<decrease-intensity>…</decrease-intensity> | Gradually decreasing intensity |
<higher-pitch>…</higher-pitch> | Raised pitch |
<lower-pitch>…</lower-pitch> | Lowered pitch |
<slow>…</slow> | Slower delivery |
<fast>…</fast> | Faster delivery |
<sing-song>…</sing-song> | Melodic, sing-song pattern |
<singing>…</singing> | Full singing mode |
<laugh-speak>…</laugh-speak> | Spoken with laughter mixed in |
<emphasis>…</emphasis> | Stressed emphasis on words |
Example:
"Welcome back! [pause] <whisper>Just between us</whisper>, the sale ends tonight. [laugh] Don't miss it!"
Rate Limits
| Limit | Value |
|---|---|
| Requests per minute (RPM) | 3,000 |
| Requests per second (RPS) | 50 |
| Concurrent sessions per team | 100 |
| Maximum characters per request | 15,000 |
Pricing
Pricing is based on the number of characters in the input text field.
SKU
| SKU | Description | Unit Price |
|---|---|---|
sku_per_1k_chars | Per 1,000 characters synthesized | $0.015 |
Formula
cost = countChars(text) / 1000 × sku_per_1k_chars
Where countChars(text) counts the total number of characters in the input text (including spaces and punctuation, excluding inline speech tags consumed by the model).
The effective rate is $15.00 per 1,000,000 characters.
Examples
| Input Length | Characters | Cost |
|---|---|---|
| Short sentence | 100 chars | $0.0015 |
| One paragraph | 500 chars | $0.0075 |
| Article excerpt | 2,000 chars | $0.0300 |
| Maximum single request | 15,000 chars | $0.2250 |
Custom cloned voices are billed at the same per-character rate as built-in voices — there is no additional charge for using a
voice_idfrom the Voice Library.
Enterprise Features
- SOC 2 Type II certified
- HIPAA eligibility with Business Associate Agreement (BAA)
- GDPR compliance with data residency options
- SAML SSO and role-based access control (RBAC)
- Multi-region infrastructure with custom SLAs available
- No data retention — audio is never stored or used for model training
