ElevenLabs has released two speech-to-text models, Scribe v3 for pre-recorded audio and Scribe v3 Live for low-latency real-time applications, both designed for messy audio such as overlapping speech and background noise. Both are available through the ElevenAPI and the ElevenAgents platform. According to the company, Scribe v3 cuts word errors by about 25 percent from its predecessor across 102 languages, and Scribe v3 Live adds real-time speaker diarization.
Independent benchmarks from Artificial Analysis place Scribe v3 among the most accurate transcription models on the market. It is slower and pricier than models from Microsoft and xAI, which the benchmark shows running many times faster at marginally higher error rates and at roughly half the cost. The benchmark price for Scribe v3 is $3.33 per 1,000 minutes.
For voice agents, Scribe v3 Live analyzes conversational context alongside the audio signal to judge when a speaker has finished, rather than relying on rigid pause timers that can cut callers off. Developers can preload thousands of specialized terms, set formatting rules for numbers and dates, and use a multilingual mode for mid-sentence language switches. Enterprise customers can get isolated servers in the EU and a zero-retention mode.
Why it matters if you build agents
Teams building voice agents can compare context-based end-of-turn detection with their current pause-timer setup and measure the real-time latency in their own call flows. Those processing large volumes of audio cheaply should weigh the speed and price gap, since the accuracy lead comes from benchmarks run only on recorded files.
Get stories like this every weekday
One short email with the day's agent tooling news.