What Is Grok Voice Transcribe 2.0? $0.10/Hour

Inside Grok Voice Transcribe 2.0: $0.10/Hour Batch Transcription

Grok Voice Transcribe 2.0 is a proprietary multilingual speech-to-text model from SpaceXAI that transcribes uploaded audio and live streams through the Grok Voice API, priced at $0.10 per audio hour for batch work and $0.20 per audio hour for streaming. Released on September 18, 2026, it is built on the audio foundation behind Grok Voice and supports speaker diarization, multichannel transcription, key-term biasing, text formatting, filler-word removal, and smart turn detection. SpaceXAI says it is twice as accurate as version 1.0 and ranks first among 32 streaming models on the Artificial Analysis leaderboard.

Source: @nima_owji

Key Takeaways

• Grok Voice Transcribe 2.0 is a hosted speech-to-text API, not a general chat model or an open-weight release.
• Batch transcription costs $0.10 per audio hour, while real-time streaming costs $0.20 per audio hour.
• SpaceXAI reports a short-phrase word error rate reduction from 20.6% to 6.8% across 19 languages.
• The model handles batch files, URLs, and real-time streams, with automatic language detection and language switching.
• Its developer features include Speaker Diarization, Multichannel Transcription, Key Term Biasing, and Smart Turn Detection.
• The #1 leaderboard claim is benchmark-specific; community reporting says Grok can trail other models on a separate voice-agent test.

What Is Grok Voice Transcribe 2.0?

Grok Voice Transcribe 2.0 is SpaceXAI's second-generation automatic speech recognition model. Its job is to convert spoken audio into text for applications such as voice agents, call analysis, video transcripts, dictation, and multilingual commands.

The model is available through the Grok Voice API in two operating modes. Batch Transcription processes recorded files or URLs after upload. Streaming Transcription processes live audio and returns text during the conversation. The official voice documentation lists speech-to-text pricing of $0.10 per hour for batch requests and $0.20 per hour for streaming requests.

SpaceXAI positions the model for difficult audio rather than only clean studio recordings. The release materials identify telephone audio, competing speakers, local accents, spoken credentials, and short voice commands as important evaluation conditions. The model is also described as being trained on live, noisy, multilingual audio and refined with post-training.

The release is part of the broader Grok developer ecosystem. Readers tracking that ecosystem can also see the earlier analysis of Grok 4.3 and the Grok Build surge.

Grok Voice Transcribe 2.0 at a Glance

Specification	Details
Developer	SpaceXAI, also referenced as xAI in the API documentation
Type	Proprietary speech-to-text model
Modality	Audio input to text output
Processing modes	Batch and real-time streaming
Languages	Dozens of languages; exact complete list not yet confirmed
Context window	Not yet confirmed
Batch pricing	$0.10 per audio hour
Streaming pricing	$0.20 per audio hour
Availability	Live in the Grok Voice API as of September 18, 2026
Model weights	Not published
License	Proprietary hosted API
Multichannel support	Up to 8 audio channels
Key-term biasing	Up to 100 domain terms per request
Default-model status	Not yet consistently confirmed
Self-hosting	Not available based on the release material

How Grok Voice Transcribe 2.0 Works and What Makes It Different

The published technical description gives a model lineage rather than an architecture diagram. Grok Voice Transcribe 2.0 is built on the audio foundation model behind Grok Voice, which SpaceXAI says already supports customer-support calls, video narration, and voice agents in physical products.

The main design emphasis is robustness across conditions. A short command provides little linguistic context, while an 8-kilohertz telephone recording removes much of the audio detail available in higher-quality speech. SpaceXAI's evaluation sets therefore separate telephony, conversational speech, spoken credentials, and short multilingual phrases.

The most specific reported improvement is multilingual short-phrase recognition. SpaceXAI reports that word error rate fell from 20.6% with version 1.0 to 6.8% with version 2.0 across 19 languages. That is a reduction of 13.8 percentage points, although the test is vendor-reported and has not been independently reproduced in the reporting so far.

SpaceXAI also reports conversational word error rate falling from 8.7% to 3.3%, and telephony error reaching 7.1%. These figures describe internal evaluations, not a universal score across all customer audio.

A separate public benchmark provides a useful but narrower comparison. The official release places Grok Voice Transcribe 2.0 first for accuracy among 32 streaming models on Artificial Analysis. A community report citing the leaderboard gives it about 2.7% streaming word error rate at approximately 0.49 seconds after speech ends. Public reporting also includes a later report using a denominator of 33, so the model count should be treated as time-sensitive.

Grok Voice Transcribe 2.0's strongest confirmed positioning is same-price access to a newer transcription model with broad real-world audio targets.

The feature set is unusually relevant to production pipelines:

• Speaker Diarization adds speaker labels to a transcript without a separate listed charge.
• Multichannel Transcription processes as many as 8 channels independently, useful for separated microphone or call-center tracks.
• Key Term Biasing accepts up to 100 domain terms per request, such as product names or medical vocabulary.
• Text Formatting converts numbers, dates, currencies, phone numbers, and email addresses into written forms.
• Filler-Word Removal removes words such as “um” and “uh” from the returned transcript.
• Smart Turn Detection predicts when a speaker has finished, which matters for voice-agent response timing.
• Word-level timestamps include start and end positions plus confidence scores, according to the release documentation.
• Automatic language detection can follow a language change during one recording.

What You Can Do With Grok Voice Transcribe 2.0

For recorded media, the API can turn uploaded files or URLs into searchable transcripts. That supports video narration indexing, meeting records, podcast archives, and post-call quality review. The exact batch file limits and supported formats should be checked against the current API documentation before production deployment.

For live applications, streaming transcription can feed a voice agent, call assistant, dictation tool, or accessibility interface. Smart Turn Detection can help an agent decide when to respond, while word-level confidence scores can support selective human review.

The multilingual capability is especially relevant to short commands. In-car assistants, field-service tools, and mobile applications often receive brief utterances with little context. SpaceXAI's 19-language short-phrase evaluation directly targets that class of interaction, but teams should still test their own accents, names, code words, and background noise.

Multichannel support can simplify contact-center processing. A system with separate agent and caller tracks may preserve speaker identity without relying entirely on diarization. That does not guarantee perfect labels, so a production test should measure both transcription errors and speaker-attribution errors.

Developers building adjacent voice experiences may also compare the transcription layer with a voice-output model such as ElevenLabs V3. That is a different task, since ElevenLabs V3 is a text-to-speech model rather than a speech-recognition model.

How Grok Voice Transcribe 2.0 Compares

The clearest comparison is price and benchmark scope. A community report listed GPT Live Transcribe at approximately $1.02 per hour for streaming, compared with $0.20 per hour for Grok Voice Transcribe 2.0. That price comparison is not a complete procurement analysis because quotas, latency, audio conditions, and billing definitions may differ.

Comparison point	Grok Voice Transcribe 2.0	Other reported systems
Streaming price	$0.20 per audio hour	GPT Live Transcribe reported at about $1.02 per hour
Artificial Analysis streaming result	Officially reported #1 among 32 models	Ranking changes as models and tests update
Reported streaming WER	About 2.7% in a community reading of the leaderboard	Scope and denominator require verification
Separate AA-AgentTalk result	Community report lists 3.2% final WER	Muse Voice Transcribe 2.7%; ElevenLabs Scribe v2 Realtime 2.8%
Version comparison	2.0 reported as twice as accurate as 1.0	Like-for-like independent testing is not provided

Grok Voice Transcribe 2.0 appears strongest on the price-performance evidence, but its benchmark advantage is not universal across every voice workload.

The separate AA-AgentTalk result is important because it measures a different problem from general streaming transcription. A model can perform well on final transcript accuracy while responding less effectively in a voice-agent interaction. Teams should test end-to-end task completion, turn timing, speaker labels, and correction effort rather than relying on one leaderboard.

Availability: How to Access Grok Voice Transcribe 2.0

Grok Voice Transcribe 2.0 is available through SpaceXAI's Grok Voice API for batch and streaming speech-to-text. The official documentation identifies the speech-to-text API area and lists /v1/stt as the endpoint family. Developers need an API key and should use the current vendor documentation for authentication, request formats, quotas, and regional availability.

The model is hosted rather than downloadable. No open weights have been published, and the release material does not describe a self-hosted package. The model is therefore suited to teams that can send audio to a managed API and less suited to deployments requiring local inference or strict offline processing.

Migration status needs care. Some secondary reports say existing Speech-to-Text integrations receive the improvement without code changes. Other release material says version 1.0 remains the default while version 2.0 is opt-in, with a future default switch and deprecation planned. SpaceXAI has not provided a fully consistent public migration schedule in the evidence available for this page.

What We Don't Know Yet

Several operational details remain open:

• The complete supported-language list has not been confirmed. The documentation references 25 languages for certain written-formatting behavior, while release material says the model handles dozens of languages.
• The official context window, maximum recording duration, and maximum streaming session length are not yet confirmed.
• Regional availability, rate limits, concurrency limits, retention rules, and enterprise data controls are not fully described in the available sources.
• The exact Artificial Analysis denominator is inconsistent across early reports: official material says 32 models, while a later community report says 33.
• The reported 2.7% streaming WER and 0.49-second timing need to be interpreted against the benchmark's scoring definitions.
• Independent reproduction of the 20.6%-to-6.8% short-phrase result is not yet available.
• The model's performance on meetings, accents, proper names, monetary amounts, and non-telephone audio remains under-tested in the reports reviewed.
• Claims about Loom adoption and related production deployments are secondary reports, not confirmed first-party deployment documentation.

Frequently Asked Questions

What is Grok Voice Transcribe 2.0?

Grok Voice Transcribe 2.0 is SpaceXAI's proprietary multilingual speech-to-text model for batch files and real-time audio streams. It is available through the Grok Voice API at $0.10 per audio hour for batch transcription and $0.20 per audio hour for streaming.

How much does Grok Voice Transcribe 2.0 cost?

Grok Voice Transcribe 2.0 costs $0.10 per audio hour for batch transcription and $0.20 per audio hour for streaming transcription. These rates are reported as unchanged from Grok Voice Transcribe 1.0.

Is Grok Voice Transcribe 2.0 available?

Grok Voice Transcribe 2.0 is live through SpaceXAI's Grok Voice API for batch and streaming speech-to-text. The model's regional availability, quotas, and default-model migration schedule are not fully confirmed in the release material.

Is Grok Voice Transcribe 2.0 open source?

Grok Voice Transcribe 2.0 is not open source. SpaceXAI has not published model weights, so the model is accessed as a hosted API rather than self-hosted.

What languages does Grok Voice Transcribe 2.0 support?

Grok Voice Transcribe 2.0 supports dozens of languages, automatic language detection, and language switching within one recording. SpaceXAI's short-phrase evaluation covered 19 languages, while the exact complete language list is not yet confirmed.

Grok Voice Transcribe 2.0 vs GPT Live Transcribe?

Grok Voice Transcribe 2.0 is priced at $0.20 per audio hour for streaming, while a community comparison listed GPT Live Transcribe at about $1.02 per hour. Accuracy depends on the audio and benchmark, so the available reporting does not establish a universal winner.

What to watch next is the model's final language list, a stable Artificial Analysis ranking with a consistent model count, and independent tests on meetings, telephony, accents, and voice-agent turn timing.

Try related models on Kie.ai: Gemini 3.1 Flash TTS and Gemini 2.5 Pro Preview TTS.