ElevenLabs vs Cartesia: choosing streaming TTS interfaces, interruption handling and a character budget
Teams with ElevenLabs voice assets should check the TTD integration required for v4 Turbo. Teams building real-time voice applications from scratch can start by evaluating Cartesia’s context management. Test interruption and playback with identical texts, then budget for characters and peak concurrency.
When a support assistant reads its answer while generating it, the hardest moment is often a user interrupting: the old answer must stop, and the new one must still flow naturally. Both ElevenLabs and Cartesia can provide the text-to-speech (TTS) layer. Teams with existing ElevenLabs voice assets that want v4 Turbo should first assess the cost of integrating the Text to Dialogue protocol. Teams building a single-voice real-time application from scratch, with context-based generation and cancellation, can evaluate Cartesia first. Choose based on the actual voice, interface and bill, rather than the smallest latency figure on a website.
Is the text already complete?
Consider a voice assistant answering order questions: a language model returns text in chunks, TTS produces audio, and a player handles playback. This comparison covers synthesis interfaces within the same real-time voice pipeline, excluding telephony, speech recognition, the language model and complete Agents platforms. If the answer is already complete, evaluate HTTP audio streaming first. If text must be appended continuously, focus on WebSocket input buffering and completion signals. Streaming audio output does not necessarily mean continuously appendable text input.
As ElevenLabs’ protocol comparison distinguishes two paths: /v1/text-to-speech/{voice_id}/stream-input serves supported TTS models such as Flash; v4 and v4 Turbo use /v1/text-to-dialogue/stream-input. The latter registers voices first, then receives inputs containing text and voice_id. Changing an old Flash integration’s model name to eleven_v4_turbo is therefore insufficient. V4 Turbo permits only one registered voice: the Dialogue name is no reason to design it as a multi-speaker conversation.
Cartesia’s current model documentation lists Sonic 3.6. Its WebSocket uses fields including transcript, context_id and output format, returning audio chunks identified by context. Teams that need to validate upgrades first can use a dated snapshot such as sonic-3.6-2026-08-27; sonic-3.6 follows the latest stable snapshot. Pinning the model version helps determine whether a change in the voice came from a new script or a model upgrade.
When should speech begin before the sentence is finished?
ElevenLabs’ TTS path offers chunk-length scheduling; its TTD path buffers on the server and can be triggered with flush. Cartesia’s manual flush establishes boundaries linking text and audio within a context, with corresponding audio carrying a flush_id. The documented request combines an empty transcript, the same context, continue=True and flush=True. These markers have different semantics, so migration should not be reduced to replacing an SDK name.
For example, forcing output before an order number is complete may split the digits awkwardly, while waiting for the entire answer may leave the user feeling there has been too much silence. Use identical sets of complete short sentences, long identifiers and mid-answer corrections. Record when text first becomes available, when the first audio chunk arrives and when playback actually starts, comparing median and slower requests. This article does not test both services under identical network and voice conditions, so it does not name a real-world speed winner. Vendor model latency is not the total wait before a user hears speech.
When a user interrupts, Cartesia offers a cancel request with context_id; ElevenLabs’ multi-context TTS guide closes the old context and creates a new one. For v4 Turbo, check the corresponding TTD variant rather than copying the Flash example. Both implementations still require the application to stop sending old text, clear locally queued audio and discard late data from the old context. Canceling server-side generation does not automatically retract audio already queued in the player.
Budget by characters, then check peak concurrency
Assume 1,000 answers a month, each submitting 500 billable characters: 500,000 characters, initially excluding retries and regeneration after interruption. ElevenLabs’ API price list lists v4 Turbo at USD 0.04 per thousand characters and v4 at USD 0.08. The resulting usage estimates are USD 20 and USD 40 respectively. These calculations use public unit prices, not a final bill incorporating every account-specific condition.
Cartesia’s pricing page specifies one credit per character for Sonic. Pro costs USD 5 a month with 100,000 credits; enabled overages cost USD 65 per million additional credits. Under the same assumptions, 5 + 0.4 × 65 = USD 31. Startup costs USD 49 a month with 1.25 million credits, but more allowance alone is not a reason to upgrade: Pro permits 3 concurrent TTS requests and Startup 5, so peak demand may determine the plan first. This example does not rank overall value or count Agents balances as TTS credits.
An actual budget should include text resubmitted for retries, script revisions and regeneration after interruptions, with billing checked to establish how canceled requests are charged. Add speech recognition or telephony separately if used. Text density per minute varies with language, speaking rate and pauses, so the vendors’ approximate minute allowances are not equivalent workloads. Sufficient credits also do not guarantee generation requests will avoid queues or failures at peak load.
Confirm the chosen voice can be used for this business
ElevenLabs’ commercial-use documentation ties commercial rights to paid plans and rights in the input content. Its voice-management documentation requires permission for cloning, with Professional Voice Cloning intended for the user’s own voice. Cartesia lists commercial-use licensing on Pro; its acceptable use policy requires recordings to be your own or recordings of others with their explicit consent. Buying a plan does not replace the speaker’s authorization. A company possessing an employee recording should not infer permission to use that voice indefinitely for customer support.
Finally, have both services read the same authorized business texts, then compare identifier pronunciation in the target language, sentence transitions and playback stopping after interruptions. If you already have a suitable ElevenLabs voice and are willing to adapt to TTD, try v4 Turbo first. If context cancellation, model snapshots and a small-scale character budget matter more, try Cartesia first. If nobody on the team owns playback queues, interruption handling and billing monitoring, switching TTS vendors will not resolve that work.
Content note:This article is based on publicly available information and is intended as a reference for evaluating tools. Features, pricing, and related policies may change at any time; please refer to the latest official information.
Original source:elevenlabs.io