ElevenLabs has released two new speech-to-text models: Scribe v3 processes recorded audio, and Scribe v3 Live is oriented to real-time scenarios with lower latency.
The official improvements focus on difficult recordings: multiple people talking at the same time, interrupting each other, and background noise. The word error rate has dropped by about 25% compared to the previous generation, and it supports 102 languages, with significant improvement in German. Speaker segmentation is more stable in large group conversations, and the Live version supports real-time segmentation of speakers for the first time.
In the third-party Artificial Analysis test, Scribe v3 has the highest accuracy, ranking ahead of Microsoft, Mistral, xAI, Google and OpenAI, and only losing to a preview model from Alibaba.
The shortcomings are also listed in the same test: it is not fast and it is not cheap. The base price is $3.33 per 1,000 minutes, with faster and cheaper options from Microsoft and xAI. Moreover, this test only measured the error rate of recording transcription, and did not cover speaker segmentation, small languages, and real-time performance, which ElevenLabs considers to be its strengths.
The Live version has a practical change for the voice customer service team: the way to determine when a user has finished speaking no longer relies solely on fixed pause timing, but instead combines context and audio signals. Read a string of mobile phone numbers slowly and pause in the middle, so that the system will not rob you of the call. Developers can also preload industry terms and format digital dates.
Enterprise customers can choose EU dedicated servers and zero data retention mode. If your business involves accuracy-sensitive scenarios such as meeting transcription and customer service quality inspection, it is worth running your own recordings for comparison; if it is a large-volume ordinary transcription, price may be the deciding factor.