ElevenLabs Releases v4 and Turbo Voice Models
ElevenLabs has launched Eleven v4 and the low-latency version Eleven v4 Turbo, positioning them as the most emotionally expressive text-to-speech models available. They also opened ElevenAgents, ElevenCreative, and ElevenAPI.
v4 incorporates performance into scripts: it can embed emotions, pauses, laughter, whispers, actions, and environmental sounds in the dialogue, such as laughs, said angrily in French accent, light rain, phone buzzing, and can specify how a line should be delivered using natural language. In blind tests, about 75% of listeners preferred its expressiveness and naturalness, with the Artificial Analysis voice ranking it first, scoring an Elo of approximately 1319, higher than Cartesia Sonic 3.6's 1276 and Gemini 3.8 Flash TTS's 1267.
For cloning, Instant Voice Clone requires about 10 seconds of recording to create a voice; Professional Voice Clone, which was previously unavailable in v3, has returned in v4 and covers the full emotional range. Context stitching connects the rhythm and state of different generated segments, while speaker stability ensures that the same character maintains their voice across narration, dialogue, and reworked segments. The single generation limit has been raised to 10,000 characters, up from 5,000 in v3.
The same script can accommodate multiple speakers, dialogue relationships, and sound effects, with the model generating responses based on the entire context rather than piecing together isolated sentences. Pronunciation is locked using IPA and pronunciation dictionaries for names, abbreviations, and technical terms. The number of languages has expanded from over 70 in v3 to over 90, with notable improvements in Japanese, Brazilian Portuguese, Mandarin, and Cantonese; demonstrations have also included Catalan narration with sound effect labels.
Turbo compresses the same expressiveness into real-time conversations: median reasoning latency is about 100ms, median first sound output is about 150ms, supporting bidirectional streaming, allowing it to start speaking while the large model is still generating words. Official comparisons show Cartesia Sonic 3.6 at about 262ms and OpenAI GPT-4o mini TTS at about 814ms. Designed scenarios include customer service with medical terminology, fast-talking characters with slang in games, and varying tones during upgrades, confrontations, and waiting. Enterprise calls now account for more than half of its business.
In terms of market mechanisms, buyers include companies looking to launch voice agents and content teams producing audiobooks, advertisements, and voiceovers; sellers are voice model vendors that separate "speed" and "human-like" into two products. The event-driven approach is clear: reducing latency below human conversational pauses is necessary to sell emotional models into call centers. The funding flow for API is charged per character, with promotional pricing for v4 at about $22 per million characters and Turbo at about $11, with list prices around $80 and $40 respectively. Beneficiaries include ElevenLabs' agents and creative suites; those under pressure include Cartesia, Google Gemini TTS, OpenAI real-time voice, and outsourced seats still relying on monotone synthesis for customer service.
On a supplementary level, v4 removes the Style and Speed sliders and does not support SSML, with control shifting to stability, similarity, and sound effect labels. Cloning requires voice owner verification and consent, and generated audio carries its voice classifier tags. The model has entered a free tier, with paid plans starting at about $6 per month.
Source: Public Information
ABAB AI Insight
ElevenLabs, led by Mati Staniszewski and Piotr Dabkowski, has consistently aimed to first create a "human-like" creative tool before integrating the same voice into enterprise calls. v3 addressed label acting, while v4 aims to solve long text performance, short sample cloning, and the reintroduction of PVC. Following a warm-up in Warsaw, they launched both models directly, indicating they no longer view quality and latency as part of the same product line but instead sell creation and production calls separately.
The capital movement is very specific: on one hand, using 10-second cloning and over 90 languages to reduce content switching costs, and on the other, using 100ms-level reasoning to embed voice agents into LLM cycles. With enterprise calls accounting for over half of revenue, the next dollar will not come from influencer voiceovers but from medical, customer service, and gaming conversations billed by the minute. Promotions have slashed API prices to about a third of list prices, aiming to embed model IDs into customer code before Cartesia, Gemini, and OpenAI can compete on initial package latency.
This is not about creating another Siri. It is more akin to NVIDIA splitting the same architecture into training and inference cards: v4 is for production, Turbo is for conversation. The industry phase has shifted from "able to speak" to "able to perform a scene and catch the next line." Those who only sell monotone customer service voices are competing for the same downward channel as outsourced seats.
The structural judgment belongs to technological substitution. The mechanism is: when scripts can directly dictate laughter, pauses, noise, and rain sounds, and initial package latency is lower than human breathing, voice becomes no longer a subordinate waveform of model output but a choreographed performance layer. The billing units for seats, voiceover studios, and localization studios have shifted from hourly human voices to characters and milliseconds. Substitution occurs because control has shifted from those behind the microphone to those writing the labels.
ABAB News · Cognitive Laws
Latency below breathing allows machines to enter conversation
After 10-second cloning, voice becomes a replicable asset
Models that can perform will replace those that can only read scripts