Introducing Grok Voice Transcribe 2.0

Today we’re releasing Grok Voice Transcribe 2.0, our newest speech-to-text mannequin. Across our real-world evaluations, Grok Voice Transcribe 2.0 is likely one of the most correct transcription fashions out there at the moment and twice as correct as Grok Voice Transcribe 1.0, on the similar worth.

Grok Voice Transcribe 2.0 is constructed on the audio basis mannequin behind Grok Voice. Grok Voice already powers tens of hundreds of customer-support calls a day, transcribes tens of millions of hours of video narration, and runs voice brokers in bodily merchandise, together with the Grok assistant in Tesla automobiles. It is educated on a singular dataset of reside, noisy, multilingual audio recorded throughout a various set of environments and refined with post-training.

The outcome is likely one of the most correct transcription fashions for speech in real-world settings.

Most transcription fashions do nicely on clear, single-speaker audio. Real-world audio is more durable: flaky cellphone traces, competing voices, native accents, and cellphone numbers or electronic mail addresses learn aloud. We constructed Grok Voice Transcribe 2.0 for the toughest audio throughout circumstances and environments.

On the general public Artificial Analysis leaderboard, Grok Voice Transcribe 2.0 ranks first for accuracy amongst 32 streaming fashions.

Internal Evaluations

In addition to public benchmarks, we measure phrase error price on 4 inner units drawn from manufacturing site visitors: telephony audio from customer-support calls, conversations with Grok, spoken credentials comparable to account codes and electronic mail addresses, and brief multilingual voice instructions. Grok Voice Transcribe 2.0 improves on Grok Voice Transcribe 1.0 throughout all 4, and on telephony it leads each mannequin we examined.

Grok Voice Transcribe 2.0 transcribes dozens of languages, detects the language routinely, and follows mid-recording switches in a single cross. Multilingual accuracy is its largest enchancment over Grok Voice Transcribe 1.0.

Short phrases comparable to in-car instructions give the mannequin little context to determine the language. On our short-phrase set, phrase error price drops from 20.6% to six.8%.

Grok Voice Transcribe 2.0 helps superior configuration and controls. Existing Speech-to-Text API integrations get the accuracy enchancment with no code modifications:

• Batch and streaming. Transcribe recorded information and URLs, or transcribe an audio stream in actual time.
• Word-level timestamps. Each phrase has exact begin/finish instances and confidence scores.
• Speaker diarization. Label every speaker within the transcript, at no further value.
• Multichannel transcription. Transcribe as much as 8 channels independently.
• Key time period biasing. Pass as much as 100 area phrases per request, comparable to product names or medical vocabulary.
• Text formatting. Numbers, dates, currencies, cellphone numbers, and electronic mail addresses are returned in written type.
• Filler phrase elimination. Omit fillers comparable to “um” and “uh” from the transcript.
• Smart flip detection. Detect the top of a speaker’s flip for voice brokers.

Atlassian Loom is extensively used to report and share display recordings. Atlassian discovered Grok Voice Transcribe 2.0 extra correct than its present answer and now makes use of it to transcribe each video. Accurate transcripts open new AI workflows: report an motion plan in Loom, pipe the transcript into Cursor, and it makes the code updates immediately.

Grok Voice Transcribe 2.0 pricing is equivalent to Grok Voice Transcribe 1.0. Batch transcription stays $0.10 per hour of audio and streaming $0.20 per hour, with diarization, timestamps, and key phrases included.

Grok Voice Transcribe 2.0 will quickly be the default within the Speech-to-Text API, and Grok Voice Transcribe 1.0 might be deprecated within the coming weeks. To keep on it throughout the transition, pin grok-voice-transcribe-1.0.