Transcription Services: A New Era
Transcription now costs 10 cents an hour, and the key differentiator is no longer the model but whose conversations you get to train on.
SpaceXAI (explore this topic) (spacexai on 164news.com) claims its new transcriber is twice as accurate as its previous version at the same price. Its evaluation sets include production traffic, such as spoken account codes, phone numbers, email addresses, and more.
Release and Pricing
Grok Voice Transcribe 2.0 was released on September 18th, priced at $0.10 per hour for batch work and $0.20 per hour for streaming, unchanged from the previous version. The company, now operating under the name SpaceXAI since July after acquiring xAI in February, makes these claims.
Accuracy and Evaluation
SpaceXAI states its model is twice as accurate as its predecessor. However, it compares itself to other vendors' models, not to others in the field generally. It ranks first for accuracy among 32 streaming models on the public Artificial Analysis leaderboard but emphasizes that it's one of the most accurate rather than the absolute leader.
Their internal charts compare against ElevenLabs Scribe v2 and Deepgram Nova-3, which are vendor-run evaluations using vendor-chosen comparators, a common practice.
The Market Shift
The price remains constant while claiming doubled accuracy is a market statement, reflecting transcription's commoditization. Features like speaker labeling, word-level timestamps, and key term biasing were once billable but are now included for free.
OpenAI has recently released new voice API models, and Microsoft benchmarked its in-house transcription model against Whisper, Gemini, and ElevenLabs across 25 languages. The competition is fierce.
The Unique Dataset
SpaceXAI trains its model on a unique dataset of live, noisy, multilingual audio recorded from diverse environments. This includes customer support calls, video narration transcriptions, and voice commands in Tesla vehicles. They claim the model excels in telephony, accents, crosstalk, and in-car commands due to this real-world data.
Evaluation Sets
The evaluation sets include four internal datasets: customer support telephony, conversations with Grok, short multilingual voice commands, and spoken credentials (account codes, phone numbers, email addresses). Notably, the announcement does not specify how this data is obtained or consented to, raising potential privacy concerns.
Privacy Considerations
In Europe, voice recordings are considered personal data, and using customer support calls for training without explicit consent from all parties could raise legal issues.