SpaceXAI Releases Grok Voice Transcribe 2.0 With 2x Accuracy

SpaceXAI has released Grok Voice Transcribe 2.0, its latest speech-to-text (STT) model, which the development team claims offers twice the accuracy of Grok Voice Transcribe 1.0 while maintaining the same pricing structure. This new model is specifically engineered to handle challenging audio environments, including noisy phone lines, conversations with multiple speakers, regional accents, and spoken credentials. Grok Voice Transcribe 2.0 is available for deployment as a hosted API, accessible today under the model ID grok-voice-transcribe-2.0. SpaceXAI has not released open weights for the model, meaning self-hosting is not an option for users. The underlying technology for Grok Voice Transcribe 2.0 is built upon the audio foundation model that powers Grok Voice. According to SpaceXAI, Grok Voice is already processing tens of thousands of customer support calls daily, transcribing millions of hours of video narration, and running the Grok assistant within Tesla vehicles. The training data for the model comprises live, noisy, and multilingual audio recordings captured across a variety of environments, which were subsequently refined through post-training processes by the SpaceXAI team. SpaceXAI reports that Grok Voice Transcribe 2.0 has achieved a first-place accuracy ranking among 32 streaming models on the public Artificial Analysis leaderboard. This benchmark, known as AA-WER Streaming, utilizes approximately 8 hours of audio and assigns weights to AA-AgentTalk (50%), VoxPopuli (25%), and Earnings22 (25%). The methodology for this benchmark is detailed separately. In addition to the public leaderboard, SpaceXAI has also measured the word error rate (WER) on four internal datasets derived from their production traffic. These internal sets include Telephony (8 kHz) for English customer support calls, Conversational for English conversations with Grok, Credentials for spoken phone numbers, emails, and addresses in English, and Short phrases for voice-assistant utterances across 19 languages. SpaceXAI states that version 2.0 demonstrates improvements over version 1.0 across all four of these internal datasets. Specifically, on telephony data, SpaceXAI claims its model outperforms every other model tested by the company. These internal performance metrics are vendor-reported and have not been independently reproduced by third parties. The model supports transcription in dozens of languages and features automatic language detection. It can also seamlessly switch between languages during transcription. The pricing for Grok Voice Transcribe 2.0 is set at $0.10 per hour, a rate consistent with its predecessor.