Grok Voice Transcribe 2.0 is xAI’s latest speech-to-text model, priced at $0.10 per hour for batch transcription and $0.20 per hour for streaming. xAI says the new model is twice as accurate as Grok Voice Transcribe 1.0 across its real-world evaluations, while keeping the same price.
Speaker diarization, word-level timestamps and key-term controls are included, and the API supports both recorded audio and real-time streaming. xAI’s official announcement provides the launch details.
The more important marketing implication is not cheaper transcription by itself. Sales calls, support conversations, webinars, product demos and voice-agent interactions can now become structured, searchable data at very low processing cost. That makes the release relevant to marketing and RevOps teams even though the product itself is infrastructure rather than a marketing application.
It also adds another layer to the voice stack xAI has been building. We previously covered its Grok Voice Agent Builder, which is aimed at creating production voice agents. Better transcription makes those systems more useful only when the surrounding workflow can turn the transcript into routing, analysis or action.
What changed in Grok Voice Transcribe 2.0
This is more than a model swap.
Grok Voice Transcribe 2.0 supports recorded audio through a REST API and real-time audio through WebSocket streaming. xAI’s current Speech to Text documentation now lists version 2.0 as the default when developers omit the model parameter, while version 1.0 can still be pinned. That is a post-launch change: xAI’s September 17 release notes still listed version 1.0 as the default when 2.0 first became available. xAI’s current speech-to-text documentation reflects the latest configuration.
The feature set is unusually practical for business audio:
| Capability | What is supported | Why businesses should care |
| Batch transcription | Files or audio URLs, with files up to 500 MB | Useful for recorded calls, webinars, interviews and video archives |
| Real-time transcription | Streaming audio through WebSocket | Relevant to live voice agents and call-assistance tools |
| Speaker diarization | Identifies different speakers | Makes call analysis easier without manually separating speakers |
| Multichannel audio | Up to 8 channels | Useful when agent and customer audio are recorded separately |
| Key-term biasing | Up to 100 terms per request | Helps with product names, technical terms and industry vocabulary |
| Word-level timestamps | Start and end times plus confidence data | Useful for search, QA, clips and conversation analysis |
| Text formatting | Numbers, dates, currencies, emails and phone numbers | Reduces transcript cleanup before data enters another system |
The model also supports multilingual transcription, automatically detects languages and can follow language changes within the same recording. xAI says the model can transcribe dozens of languages, while the current API documentation lists 25 language codes for number, currency and unit formatting.
For the broader xAI voice stack, this follows Grok Voice Think Fast 2.0, which focused more heavily on speech-to-speech interaction, response timing and agent performance.
The accuracy claim is strong, but “2x” needs context
xAI says the new model is twice as accurate as version 1.0 across its real-world evaluations, but those figures come from vendor-run tests.
The company tested customer-support telephony, conversations with Grok, spoken credentials and short multilingual commands. One of the largest improvements came from the multilingual short-phrase test, where word error rate fell from 20.6% to 6.8% across 19 languages. Lower word error rate is better.
Independent benchmarking adds useful context. Artificial Analysis’ streaming benchmark is currently reported at about 2.7% final word error rate for Grok Voice Transcribe 2.0 versus 3.9% for version 1.0, with the newer model taking roughly 0.49 seconds to return the final transcript after speech ends versus about 0.37 seconds for version 1.0. xAI also cites the public Artificial Analysis leaderboard in its launch announcement. Artificial Analysis tracks the current streaming speech-to-text benchmark.
The result is a useful trade-off to test rather than a universal win: version 2.0 appears more accurate on this benchmark, while version 1.0 reached the final transcript faster.
For Grok Voice Transcribe 2.0, that trade-off matters more in a live voice agent than in a recorded webinar. Half a second is irrelevant when processing yesterday’s sales calls. Inside a spoken conversation, every fraction of a second stacks with reasoning, tool calls and text-to-speech latency.
The same caution applies when comparing this release with Gemini 3.5 Transcribe, Microsoft MAI-Transcribe-2-Streaming, or GPT-Live and OpenAI’s voice direction. Different benchmarks measure different audio, configurations and product architectures.
One leaderboard position is useful evidence, not a universal verdict.
The $0.10 price matters more than another leaderboard win
The economics may be the more significant part of this announcement.
Batch transcription remains $0.10 per audio hour and streaming remains $0.20 per hour, with diarization, timestamps and key-term biasing included.
That pushes speech-to-text AI further away from being a specialist feature and closer to basic infrastructure.
Historically, teams had reasons to be selective about what they transcribed. Maybe only sales calls were processed. Support conversations stayed somewhere else. Webinars became a recording nobody reopened. Customer interviews disappeared into folders with names such as “final_v3_revised.”
At these prices, the more useful question becomes: which conversations should be searchable?
Cheap transcription does not create customer insight by itself. It simply removes one cost barrier between the conversation and whatever comes next.
How marketers can turn transcripts into something useful
For marketing teams, Grok Voice Transcribe 2.0 becomes interesting when transcription feeds another workflow.
1. Turn sales calls into customer-language data
A transcript can capture pricing objections, competitor mentions, product terminology, buying criteria and recurring questions.
One call proves almost nothing. A few hundred conversations can reveal patterns.
That data can feed positioning research, landing-page copy, paid-search messaging and sales enablement. It also fits the wider shift toward AI marketing automation, where the useful automation happens after clean inputs reach the right system.
2. Give voice agents cleaner inputs
Voice agents cannot reason around words they heard incorrectly.
Better real-time transcription, key-term biasing and smart turn detection could reduce errors before a reasoning model or workflow engine takes over.
That connection matters as enterprise systems such as OpenAI Presence push voice and chat agents deeper into support, sales and operational workflows.
The transcript is still only one layer. Routing, permissions, CRM updates and human escalation remain separate problems.
3. Repurpose spoken content with less cleanup
Webinars, podcasts, product demos, customer interviews and Loom recordings already contain useful source material.
More accurate transcripts can shorten the messy first step between recording something and turning it into searchable notes, briefs, FAQs, clips or written content.
xAI says Atlassian found the model more accurate than Loom’s previous transcription solution and is now using it for Loom video transcription. That is a customer-reported result, not a controlled independent benchmark.
4. Build better feedback loops around customer conversations
Transcription becomes more valuable when it connects with CRM, analytics and agent workflows.
An always-on system such as Grok Bot could eventually help teams combine conversation data with other operational signals.
The useful outcome is not “AI summarized the call.” It is identifying repeated objections, routing them to the right owner and measuring whether the response changed conversion behaviour.
That is where voice AI starts affecting GTM execution rather than producing prettier meeting notes.
The cheap API still comes with a data question
There is one part marketing and RevOps teams should check before uploading thousands of customer conversations: retention.
xAI’s current API security documentation says API inputs and outputs are not used for model training without explicit permission, but standard API requests and responses are retained for 30 days for abuse and misuse auditing.
Zero Data Retention is available for teams that need stricter handling, but it disables features that depend on stored content, including per-API-key request logging, stateful Responses, Files and Collections, Batch processing and other storage-dependent workflows. xAI’s security documentation explains the current retention options and trade-offs.
Call transcripts can contain names, phone numbers, commercial information and other sensitive data. A $0.10 transcription price does not remove the need for consent, access controls, retention policies and legal review.
The transcription model may be cheap, but poor data governance can still be expensive. Teams should decide what audio can be sent, who can access transcripts and how long the resulting data should be retained before scaling the workflow.
What marketing teams should watch next
Grok Voice Transcribe 2.0 is now listed as the default transcription model in xAI’s current API documentation. The launch announcement said version 1.0 would be deprecated in the coming weeks, so teams still pinning the older model should plan and test their migration deliberately.
Teams using the older model should test their own audio rather than assuming benchmark gains will transfer perfectly.
Three areas deserve attention next:
- Accuracy on actual customer audio. Test accents, poor microphones, phone lines, interruptions and your own product vocabulary.
- Latency inside full voice workflows. Transcription is only one delay. Measure the complete time from the customer finishing a sentence to the agent responding.
- How transcripts connect to systems of record. The real business value appears when useful information reaches CRM, support, product and content teams without creating another disconnected dataset.
The model ranking will change. Those workflow questions will remain.
OneMetrik Takeaway
Grok Voice Transcribe 2.0 matters less because xAI can claim another benchmark lead and more because high-quality transcription is getting cheap enough to become an invisible layer inside everyday business workflows.
For marketers, that makes customer conversations more usable as data.
But transcription is not customer research. A transcript still needs aggregation, analysis, context and judgment before a recurring phrase becomes a campaign insight.
The sensible test is small: take one meaningful source of customer audio, compare the transcript against your current setup, measure correction effort and then see whether the resulting data improves an actual workflow. Scale only when the transcript reliably improves downstream work. If the only outcome is a cheaper pile of transcripts, the model may have worked, but the marketing system has not.