Real-Time AI Voice APIs Compared: A Buyer Guide

A sensible first trial for a low-latency streaming build is Cartesia, especially when partial LLM output, persistent WebSocket sessions, telephony-ready audio, or a path into a managed agent runtime matters. That is a starting route, not a total winner. ElevenLabs can be the better platform trial when the same purchase must cover agents, cloning, and broader audio creation; Fish Audio and MiniMax Audio are credible speech-layer routes for teams that own orchestration; Resemble AI becomes more relevant when a first-party agent control plane, voice governance, or on-premises deployment is central.

The important decision is therefore not “which vendor has the smallest latency number?” It is which layer the product needs the vendor to own, which official meter matches the workload, and whether the published limits and regional controls survive a production test.

Start with the job, not the vendor

Real-time TTS accepts text and starts returning playable speech before the full synthesis job is finished. It may use a streaming HTTP response, server-sent events, or a persistent WebSocket. This is the right layer when the product already owns speech recognition, turn detection, the LLM, tools, call state, and monitoring.

A voice-agent platform owns more of the live conversation. It can add speech input, interruptions, turn-taking, tools, knowledge retrieval, telephony, logs, evaluations, and deployment controls. Cartesia Line, ElevenAgents, and Resemble Agents are documented first-party agent routes. Fish Audio is positioned more safely as a speech layer for frameworks such as LiveKit, while MiniMax's current speech docs clearly support TTS inside an agent stack but do not establish a current, fully documented hosted-agent product.

Voice cloning is another purchase. An instant clone or professional clone creates a reusable voice identity; it does not supply conversation orchestration. General voice generation is broader still: long-form narration, voice design, dubbing, sound effects, voice changing, music, or studio editing may matter even when millisecond-level response time does not.

Keep these four jobs separate in the requirements. Otherwise a broad creator platform can look like the best agent API, or a fast TTS endpoint can be mistaken for a complete calling system.

Decision table

The table routes a trial; it does not turn unlike products into a total ranking. A buyer who only needs speech output should not pay for agent ownership by default, while a buyer who needs calls, tools, and observability should not choose on TTS unit price alone.

Why Cartesia is the contextual starting route

Cartesia's strongest case is the streaming path itself. Its persistent WebSocket accepts partial input through continuations, keeps prosody across chunks, and can multiplex multiple active context IDs on one connection. The bytes and SSE endpoints also stream output, but they take a complete transcript and do not offer the same continuation behavior. That makes WebSocket the clearest Cartesia trial for an LLM that is still producing text while speech begins.

The concurrency model is also concrete enough to test. Each active WebSocket context consumes TTS concurrency, while more messages inside the same context are processed sequentially. WebSocket connections may reach ten times the TTS concurrency allowance, idle connections close after five minutes, and exceeding active-context limits returns a rate-limit response. Line has its own concurrent-call pool, so raw TTS capacity must not be used as a proxy for hosted-agent capacity.

Line creates a credible expansion path without making raw Sonic TTS and hosted agents the same product. It adds audio orchestration, turn-taking, deployment, logs, evaluations, and phone or custom WebSocket entry. Its per-minute meter is separate from model credits. A buyer should therefore run one TTS-only budget and another Line budget rather than assuming the monthly credit allowance pays for the full call stack.

Cartesia is not the default for every voice job. The current Sonic 3.5 model page does not provide a reproducible numeric end-to-end benchmark, official pages conflict on the exact self-serve floor for Instant Voice Clone, and country-level cloud pinning is not documented publicly. Pro Voice Clone also has a model-pinning boundary: training costs one million credits, output uses about 1.5 credits per character, and retraining against a new base model incurs another training charge.

When broader platform fit beats the starting route

ElevenLabs

Choose ElevenLabs as the first platform trial when real-time TTS is only one part of the roadmap. ElevenAgents is a complete documented agent layer with visual workflows, tools, knowledge, custom or supported LLMs, web and mobile SDKs, telephony, experiments, monitoring, and analytics. The same vendor also offers Instant and Professional Voice Cloning, Voice Design, a voice library, dubbing, voice changing, speech-to-text, sound effects, and music.

Its commercial model is correspondingly layered. TTS character rates vary by model family; agent calls use minutes; burst calls cost more; telephony and LLM usage are additional; and adding pay-as-you-go funds does not raise plan-level concurrency or feature limits. This route fits breadth, but the buyer must map each production component to the correct meter.

ElevenLabs also publishes the clearest self-serve geographic latency guidance in this group. Flash over WebSocket is routed through the USA, Netherlands, or Singapore, and the response can expose the serving region. Enterprise isolated environments are separate deployments in the EU, India, and Singapore. Routing near a user is not the same promise as data residency, and standard customer data remains hosted in the United States.

Fish Audio

Choose Fish Audio when the product already owns the agent runtime and needs a direct synthesis and cloning layer. The bidirectional MessagePack WebSocket accepts incremental text, an explicit flush action, and returns audio chunks. Fish documents integrations with agent frameworks, but that evidence supports Fish as the voice layer rather than a hosted replacement for turn detection, tools, conversation state, and operations.

The direct price is easy to quote but not automatically easy to compare. Fish charges by UTF-8 bytes, so languages with different byte density can produce different effective costs for similar visible text. Concurrency increases after cumulative prepaid thresholds, and the official page explicitly avoids promising a fixed requests-per-minute conversion. Measure actual request duration before estimating throughput.

Current official pages also require caution around the newest S2.1 route. The HTTP reference includes S2.1, while the WebSocket reference names S1 and S2 Pro; published latency figures use different modes and definitions. Hosted data defaults to the United States, inference uses US and Asia-Pacific edges with Tokyo named, and stricter region control is routed through enterprise zero-retention terms or self-hosting.

MiniMax Audio

Choose MiniMax Audio when the workload mixes real-time multilingual synthesis with long-form output, rapid cloning, or generated voices. The synchronous route supports HTTP and WebSocket streaming with a 10,000-character request ceiling, while the asynchronous route expands to one million characters. That is a useful split between live output and long-form generation rather than one endpoint forced to serve both.

MiniMax publishes a vendor-reported under-250-millisecond end-to-end figure for Speech 2.6 in conversational scenarios, not a universal time-to-first-audio guarantee for current Speech 2.8 traffic. Its public limits are expressed as requests per minute, including 60 for TTS and voice cloning and 20 for Voice Design; the audio-subscription tiers publish a separate 10, 50, 200, 500, and 800 RPM ladder. No public simultaneous WebSocket-session limit or hosted-region matrix was located.

The pricing choice is also two-lane. Pay-as-you-go uses characters and per-voice charges, while subscriptions use audio points, voice slots, and tier-specific RPM. Do not invent a conversion between points and characters. The public speech docs support MiniMax as a component in an agent stack, but a buyer seeking a current first-party agent control plane should confirm that route directly rather than relying on an older Realtime API announcement.

Resemble AI

Choose Resemble AI when the requirements start with a managed agent API, voice governance, or deployment control. Its Agents API combines speech recognition, TTS, the LLM, turn-taking, tools, webhooks, dynamic variables, knowledge retrieval, and inbound or outbound telephony. Its enterprise route adds higher concurrency, service commitments, custom tuning, identity controls, and on-premises deployment.

Resemble's current model docs report best-case time to first speech of 200 milliseconds for Chatterbox-Turbo and 250 milliseconds for Chatterbox and Chatterbox Multilingual, while warning that cold starts, network latency, and load can increase those results. The persistent WebSocket is the documented lowest-latency route, but it is gated to Business and above. That plan label is not present on the current public pricing page, so entitlement needs confirmation.

Price transparency is the main pre-purchase caveat. The current live page starts Flex at no subscription charge and uses funded credits that do not expire, but it no longer publishes TTS, agent, cloning, voice-design, or voice-changing rates. Older cached amounts are not safe evidence. Treat dashboard or sales confirmation of the actual meters, plan gate, concurrency, and retention terms as a mandatory buying step.

Compare latency, scale, and region in one production test

Vendor latency numbers do not share one denominator. ElevenLabs describes Flash at roughly 75 milliseconds of model inference and explicitly excludes network and application overhead. Resemble reports best-case time to first speech. MiniMax's figure is end-to-end for a named older model. Fish publishes several first-audio figures across launch, streaming-mode, and enterprise pages. Cartesia markets low model latency, while its current Sonic 3.5 reference does not give a reproducible numeric benchmark.

A fair bake-off starts the clock when the application has enough text to send and stops at the first playable audio frame on the target device. Record the intermediate times too: connection setup, first request byte, first audio byte, decoder buffering, and full turn completion. Use the same script, voice type, audio encoding, text-chunk policy, punctuation pattern, deployment region, and warm-versus-cold connection rule.

Test interruptions and incomplete clauses, not just polished sentences. A low model figure can be erased by LLM token cadence, input buffering, network distance, audio decoding, speech recognition, end-of-turn detection, tool calls, or telephony. For hosted agents, measure the complete listen-think-speak loop instead of adding separate vendor benchmarks together.

Concurrency needs the same workload discipline. Cartesia counts active TTS contexts; ElevenLabs counts generations while audio is being produced; Fish uses concurrent in-flight requests; MiniMax publishes RPM without a public simultaneous-session figure; Resemble separates REST request rate from WebSocket session caps. Run sustained calls with realistic silence, interruptions, retries, and burst arrival instead of converting those limits into a made-up common unit.

Region routing and data control must also stay separate. ElevenLabs names global routing locations and separate enterprise isolated environments. Cartesia documents origin-based US, EU, and APAC routing but no public country pinning matrix.

Fish documents US-hosted data plus US and Asia-Pacific inference edges, with self-hosting for stricter control. MiniMax publishes no hosted-region matrix in the reviewed speech docs. Resemble says its services and personal data are hosted in the United States and offers enterprise on-premises deployment. None of those statements should be stretched into a residency promise it does not make.

Make cloning and generation a separate decision

For instant cloning, compare the required recording, consent workflow, output quality, and where the clone may be used. Cartesia takes a short clip for Instant Voice Clone and offers a separate data-heavy Pro route. ElevenLabs separates Instant from Professional Voice Cloning and gates Professional access to Creator or above. Fish supports reference-audio synthesis, stored voice models, professional cloning, and Voice Design. MiniMax accepts 10 seconds to five minutes for rapid cloning and charges when the clone is first used. Resemble documents rapid and professional routes, but API access is plan-gated.

For general generation, ask whether the team needs a developer endpoint or a human studio. ElevenLabs has the broadest documented creative-audio surface in this set. Fish combines TTS, ASR, Voice Design, and a wider web creation product. MiniMax separates live synthesis from million-character asynchronous generation. Cartesia centers Sonic speech generation and Line. Resemble combines TTS, speech-to-speech, designed voices, agents, and enterprise deployment.

Consent and commercial rights remain a human gate. A technically available clone is not proof that the buyer owns the speaker's identity, source recording, distribution rights, or permission for every use. Capture consent, intended channels, retention, revocation, and approval evidence before training or deploying a real person's voice.

Price the actual runtime

Normalize the workload before comparing rates. Count text sent for successful and retried generations, listening audio, generated audio, concurrent sessions, full agent-call minutes, telephony, LLM calls, evaluations, clone training, seats, and overages. Then map each quantity to the vendor's official unit instead of forcing credits, characters, UTF-8 bytes, audio points, seconds, and call minutes into one unsupported shortcut.

Separate prototype cost from production eligibility. A low entry price can still sit below the required concurrency, WebSocket, professional-clone, residency, or support tier. Cartesia and ElevenLabs publish detailed self-serve limits; Fish ties higher API concurrency to prepaid thresholds; MiniMax separates PAYG from subscription RPM; Resemble requires a current commercial quote for advanced voice usage.

Before committing, get written answers for every unpublished boundary: enterprise concurrency, service level, data retention, region pinning, telephony pass-through, burst charges, clone ownership, and what happens when a model version changes. The right provider is the one whose proven runtime and commercial contract fit the product—not the one with the most attractive isolated demo number.

Final selection checklist

• Choose Cartesia's WebSocket trial first when partial LLM text, warm connections, multiplexed contexts, telephony audio, or a Line expansion path defines the job.
• Branch to ElevenLabs when one platform must cover agents, several cloning routes, and broader creative audio generation.
• Branch to Fish Audio when the product owns orchestration and byte-metered TTS or a self-hosting path is attractive.
• Branch to MiniMax Audio when multilingual streaming, long-form synthesis, rapid cloning, or voice design matters and RPM-based limits fit the workload.
• Branch to Resemble AI when a first-party agent API, governance, or on-premises deployment matters enough to justify sales validation of pricing and entitlement.
• Run the same regional, transport, concurrency, interruption, and budget test on the smallest credible pair. Do not select from vendor latency labels alone.