ElevenLabs, which develops AI-powered voice technologies, has introduced its text-to-speech models Eleven v4 and Eleven v4 Turbo. The company says that Eleven v4, which features an entirely new architecture, produces more natural voices by interpreting intonation, pace, emotion, character and context; it can preserve speaker identity in long texts and change intonation based on previous lines. Multiple voice tags can be used in the same command; laughter, anger, accents and ambient sounds can be described in natural language. Improved International Phonetic Alphabet (IPA) support is also offered.
Eleven v4 can clone a voice from a 10-second recording and brings back the Professional Voice Clone feature, which was not available in Eleven v3. Both models support more than 90 languages; Eleven v3 supported approximately 70 languages. Quality improvements are targeted for Japanese, Brazilian Portuguese, Mandarin and Cantonese.
Eleven v4 Turbo was developed for call centers and voice-based AI agents. With bidirectional streaming, it starts generating early; in the company’s tests, the median time to first audio was approximately 150 milliseconds, while inference latency was measured at approximately 100 milliseconds. Eleven v4 can generate content of up to 10 thousand characters.
Why it matters
With the new architecture, maintaining speaker consistency in long-form content and managing multiple voice elements in a single command are taking the model beyond short voiceovers and into more complex production workflows. Expanding language coverage from approximately 70 to more than 90 broadens the reach and quality targets, particularly for teams using Japanese, Brazilian Portuguese, Mandarin and Cantonese. The focus on bidirectional streaming and low latency in the Turbo version shows that response time is one of the key metrics for call centers and voice-based AI agents. However, since the announced latency results are based on the company’s tests, it remains unclear whether performance will be at the same level across different infrastructures and real-world usage conditions. Offering cloning from a short sample alongside the professional cloning option also raises the question of which method users will choose for which quality requirements.
Background
ElevenLabs is not a new name in the FikirPilot archive: we published a news story mentioning the name in the last 90 days; that article was dated September 12, 2026.