xAI STT v1 API by xAI

xAI STT v1 — Speech to Text

Developer: xAI Model ID: xai/stt-v1 Release Date: April 2026

Overview

xAI STT v1 is a production-grade speech-to-text model from xAI, the company behind Grok. It transcribes audio into accurate, formatted text either in a single batch API call or in real time over a WebSocket stream. The model supports 24+ languages with automatic language detection, word-level timestamps, speaker diarization, multichannel transcription, and Inverse Text Normalization that renders numbers, currencies, and units in their written form.

xAI STT v1 is built on the same audio infrastructure that powers Grok Voice, Tesla in-vehicle assistants, and Starlink customer support — infrastructure that has been battle-tested at scale across consumer and enterprise workloads. It is engineered for demanding real-world use cases such as call-center analytics, voice agents, meeting transcription, and media captioning, where named-entity accuracy (names, account numbers, dates) and low latency matter most.

In xAI's published benchmarks the model reports a 6.9% word error rate on general audio and a 5.0% error rate on phone-call entity recognition — ahead of ElevenLabs (12.0%), Deepgram (13.5%), and AssemblyAI (21.3%) on the same entity task — while matching the leading providers (≈2.4% error) on clean video/podcast audio.

Key Capabilities

• Batch and streaming transcription — Transcribe an uploaded file or remote URL in a single request, or stream audio in real time over WebSocket for live captioning and voice agents.
• 24+ languages with auto-detection — Transcribes speech in any supported language; leave language unset to auto-detect, or set it to enable language-specific text formatting.
• Word-level timestamps — Every word is returned with precise start / end times in seconds.
• Speaker diarization — With diarize: true, each word carries a speaker_id identifying the detected speaker.
• Multichannel transcription — Transcribe up to 8 audio channels independently, returning a per-channel result set.
• Inverse Text Normalization — With text_normalization: true, spoken numbers and currency are converted to written form (e.g. "one hundred dollars" → "$100").
• Keyterm biasing — Bias the model toward up to 100 domain-specific terms (product names, jargon) of up to 50 characters each.
• Filler-word control — Remove disfluencies like "um" and "uh" by default, or retain them with filler_words: true.
• Broad format support — Auto-detects common container formats; accepts raw/headerless audio with explicit format hints.
• Privacy-first — Audio is not stored or used for model training.

Use Cases

• Call-center analytics & QA — High-accuracy entity recognition for names, account numbers, and dates in noisy phone audio.
• Real-time voice agents — Low-latency streaming transcription paired with xAI's TTS and Voice Agent APIs for end-to-end conversational AI.
• Meeting & interview transcription — Speaker-attributed transcripts via diarization with word-level timing.
• Media captioning & subtitling — Accurate transcription of podcasts, videos, and broadcasts.
• Multilingual workflows — Single API covering 24+ languages with automatic detection.
• Compliance & search — Searchable, timestamped transcripts for archives and regulated industries.

Supported Audio Formats

Container formats (auto-detected — set audio_format to auto): WAV, MP3, OGG, Opus, FLAC, AAC, MP4, M4A, MKV.

Raw / headerless formats (require audio_format plus sample_rate): PCM, µ-law (mulaw), A-law (alaw).

Technical limits:

Property	Value
Maximum file size	500 MB
Channels	Mono, stereo, or up to 8 channels
Sample rates	8000, 16000, 22050, 24000, 44100, 48000 Hz

Supported Languages

Code	Language	Code	Language
ar	Arabic	ja	Japanese
cs	Czech	ko	Korean
da	Danish	mk	Macedonian
nl	Dutch	ms	Malay
en	English	fa	Persian
fil	Filipino	pl	Polish
fr	French	pt	Portuguese
de	German	ro	Romanian
hi	Hindi	ru	Russian
id	Indonesian	es	Spanish
it	Italian	sv	Swedish
th	Thai	tr	Turkish
vi	Vietnamese

Leave language unset for automatic detection. The model transcribes any supported language regardless of this setting; specifying it together with text_normalization: true enables locale-aware formatting of numbers, currencies, and units.

Input Parameters

Parameter	Type	Default	Description
model	string	xai/stt-v1	Model identifier. Required.
audio	string	—	Publicly accessible URL of an audio file (max 500 MB) or a base64-encoded audio string. Required.
language	string	—	ISO 639-1 language code (Filipino uses fil). Leave unset for automatic detection.
text_normalization	boolean	false	Enables Inverse Text Normalization (spoken → written numbers/currency). Requires language when true.
diarize	boolean	false	Enables speaker diarization; adds speaker_id to each word.
multichannel	boolean	false	Transcribes each audio channel independently.
filler_words	boolean	false	Includes filler words ("um", "uh") when true; removed by default.
keyterm	string[]	[]	Up to 100 terms (≤50 chars each) to bias transcription toward.
audio_format	string	auto	auto for container formats; pcm, mulaw, or alaw for raw audio.
sample_rate	integer	—	Sample rate in Hz. Required for raw audio.
channels	integer	—	Number of channels (2–8). Required only for multichannel raw audio.
enable_sync_mode	boolean	false	When true, the request waits until the result is generated and uploaded, returning it directly in the response. API-only.

Response

The transcription result (stt_result) contains:

Field	Description
text	Full transcribed text.
language	Detected language name.
language_code	Detected language code (ISO 639-3).
language_probability	Confidence of the detected language (0–1).
duration	Audio duration in seconds.
words[]	Word-level segments, each with text, start, end, type, and speaker_id (when diarized).

When multichannel is enabled, stt_result additionally contains a nested channels[] array — each element holds a full per-channel transcription result (text, words[], etc.) plus a zero-based channel index.

Pricing

Price configuration path: models/xai/stt/price

Pricing for xAI STT v1 is metered by the duration of the transcribed audio (in minutes), derived from the input audio. The minute-based rate corresponds to xAI's published batch transcription pricing of **$0.10 per hour** ($0.10 ÷ 60 ≈ $0.0017 per minute).

SKU

SKU	Description	Unit Price
sku_minute	Price per minute of audio	$0.0017

Formula

cost = parseVideoDuration(audio) / 60.0 × sku_minute

Where parseVideoDuration(audio) returns the audio duration in seconds; dividing by 60.0 converts it to minutes, which is multiplied by the per-minute unit price.

Examples

Audio Length	Duration (s)	Duration (min)	Cost
Short clip	30 s	0.5	$0.00085
Phone call	5 min	5	$0.0085
Podcast episode	60 min	60	$0.1020
Long recording	120 min	120	$0.2040

The effective rate is **≈ $0.10 per hour** of audio. Real-time streaming transcription is priced higher than batch by xAI (≈$0.20 per hour); the configuration above reflects the batch per-minute rate.

Benchmarks

Task	xAI STT v1	ElevenLabs	Deepgram	AssemblyAI
Phone-call entity recognition (error rate)	5.0%	12.0%	13.5%	21.3%
Video / podcast transcription (error rate)	2.4%	2.4%	3.0%	3.2%
General audio (word error rate)	6.9%	—	—	—

Source: xAI published benchmarks (April 2026).

References

• xAI Voice / API Overview
• Speech-to-Text API Docs
• xAI Grok STT & TTS Launch Announcement
• MarkTechPost — xAI Launches Standalone STT/TTS APIs