Grok Voice Transcribe 2.0 will become the speech-to-text default at the same price

xAI released Grok Voice Transcribe 2.0 on September 18, 2026. Across its real-world evaluations, the company says the model is one of the most accurate transcription models available and twice as accurate as Grok Voice Transcribe 1.0, at the same price.

Batch transcription stays at $0.10 per hour of audio. Streaming stays at $0.20 per hour. Diarization, word timestamps, and key terms stay included in that price. Grok Voice Transcribe 2.0 will soon be the default in the Speech-to-Text API, and Grok Voice Transcribe 1.0 will be deprecated in the coming weeks. Existing integrations get the accuracy change with no code changes. To stay on Grok Voice Transcribe 1.0 through the switch, pin grok-voice-transcribe-1.0.

Public rank and four internal sets

On the public Artificial Analysis leaderboard, xAI says Grok Voice Transcribe 2.0 ranks first for accuracy among 32 streaming models. A chart in the announcement plots accuracy against price for streaming and for batch. The caption says upper and to the right is better, and the source line cites Artificial Analysis. The announcement text does not print the chart’s coordinates.

xAI also measures word error rate on four internal sets drawn from production traffic. Grok Voice Transcribe 2.0 improves on Grok Voice Transcribe 1.0 on all four. On telephony it leads every model xAI tested.

The sets are English telephony at 8 kHz from customer-support calls, English conversations with Grok, spoken credentials, and short phrases. The credentials section names account codes and email addresses. The chart label for that set names phone numbers, emails, and addresses. The short-phrase set is voice-assistant utterances in 19 languages.

The only word error rates printed in the announcement are on that short-phrase set. The rate drops from 20.6% to 6.8%. xAI says short phrases such as in-car commands give the model little context to identify the language.

Controls for the audio xAI calls hard

xAI says most transcription models do well on clean, single-speaker audio. It built Grok Voice Transcribe 2.0 for harder audio, including flaky phone lines, competing voices, local accents, and phone numbers or email addresses read aloud.

The model transcribes dozens of languages, detects the language automatically, and follows a language switch in the middle of one recording in a single pass. xAI says multilingual accuracy is its largest improvement over Grok Voice Transcribe 1.0.

The Speech-to-Text API accepts a recorded file or URL, or a live audio stream. Each word can carry a start time, an end time, and a confidence score. Speaker diarization is included at no additional cost. Multichannel requests can cover up to 8 channels, transcribed independently. A request can pass up to 100 key terms, such as product names or medical vocabulary. Numbers, dates, currencies, phone numbers, and email addresses come back in written form. The API can omit fillers such as “um” and “uh”. Smart turn detection marks the end of a speaker’s turn for voice agents.

Grok Voice Transcribe 2.0 is built on the audio foundation model behind Grok Voice. xAI says Grok Voice already powers tens of thousands of customer-support calls a day, transcribes millions of hours of video narration, and runs voice agents in physical products, including the Grok assistant in Tesla vehicles. Training used live, noisy, multilingual recordings from many environments, then post-training.

Atlassian reports a gain on Loom videos

Atlassian found Grok Voice Transcribe 2.0 more accurate than its existing solution for transcribing Loom videos, xAI says. The announcement describes a workflow that records an action plan in Loom, sends the transcript to Cursor, and has Cursor make the code updates. Sanchan Saxena, SVP of Teamwork Collection at Atlassian, says Grok handles Loom’s speech-to-text and Cursor turns that transcript into code. The announcement states that comparison and that workflow, and it stops before an error rate or a measured coding result.

Five rows cover an 8 kHz support call with no printed rate, short phrases from 20.6% to 6.8% word error, included diarization, streaming at $0.20 an hour, and a key-term list of up to 100 — Short-phrase word error falls from 20.6% to 6.8%. The telephony set, where xAI says the model leads, prints no rate.

Practical checklist

Separate the short-phrase rate from the audio you will send

Word error on the short-phrase set falls from 20.6% to 6.8%, about one third of the earlier rate. Half of 20.6% is 10.3%, and the printed rate is lower than that. Correct words, taken as 100 minus the word error rate, move from 79.4% to 93.2%. That is not twice the earlier share of correct words. “Twice as accurate” has to refer to some other aggregate, or to a definition the announcement never gives. The other three internal charts do not print their rates.

Score a sample of real customer-support calls at 8 kHz, with the line noise you actually get. Count word errors, then count errors in names, order numbers, and other fields the call is supposed to capture. xAI says this English telephony set is where Grok Voice Transcribe 2.0 leads every model it tested, and the announcement prints no rate for it.
Score short commands in each language you support, including local accents, and score a clip that switches language midway. xAI calls multilingual accuracy the largest gain over Grok Voice Transcribe 1.0. The 20.6% to 6.8% drop is on short phrases in 19 languages, where the model has little surrounding speech to identify the language. A long, clean English monologue will not show that case.
Score competing voices on one mixed channel separately from audio that already arrives on separate channels. Diarization labels speakers and is included in the hourly price. Independent transcription of up to 8 channels is a different control. Put the other vendor’s diarization charge into its hourly price before you treat $0.10 or $0.20 as the lower cost.
For a live agent, measure how long a usable partial transcript takes, and whether turn detection ends the caller’s turn too early. Streaming is priced at twice the batch rate. The announcement publishes no latency number. Its Artificial Analysis chart compares accuracy with price.
Read phone numbers, email addresses, street addresses, and account codes aloud. Compare the written form the API returns with the words that were spoken. Repeat the clip with a key-term list and without one. The list holds up to 100 terms. Use the per-word confidence score before an agent stores a credential.

Run those clips before the default changes. No code change means current Speech-to-Text API traffic moves onto Grok Voice Transcribe 2.0 when that default flips. Pin grok-voice-transcribe-1.0 to keep the current model while you compare the clips. The hourly price is the same either way.

Sources

xAI: Introducing Grok Voice Transcribe 2.0