Best Low Latency Text to Speech APIs for Voice Agents 2026

Deepgram Flux TTS is the best overall low-latency API for a voice agent in 2026 because its published as-low-as-80-ms first audio comes with native barge-in state, not just fast synthesis. Pika can generate a minute of 48 kHz speech in about 1.2 seconds, but its live API is asynchronous and explicitly not for real-time streaming, so that headline does not buy a faster phone call.

The Short Answer

Choose Deepgram Flux TTS for an English-language voice agent where callers interrupt, change their minds, and expect the agent to remember exactly what it finished saying. Its latency claim is not the smallest number on this page. Its advantage is that the streaming protocol reports both the text the caller heard and the text that was cut off when the client supplies its playback offset, which protects the conversation state after barge-in.

Choose Rime Mist v3 when transparent percentile reporting and very fast first audio matter most. Choose Inworld Realtime TTS-2 Flash when multilingual reach and paid-plan economics dominate. Choose ElevenLabs Flash v2.5 when a mature multilingual voice ecosystem matters enough to accept a higher character rate and a text-normalization chore.

Choose Mistral Voxtral TTS 3 for a low-cost pilot within its nine documented languages when complete-text requests fit the call flow and public-preview risk is acceptable.

Prices and specifications below were verified on maker pages on October 21, 2026. “Starting price” means the smallest public entry point, not the likely production bill.

The explicit decision rule is simple: interruption correctness beats a lower isolated latency number. Once two candidates handle your barge-in path correctly, compare p95 first audio through the real telephony route, then compare effective cost and voice fit. A vendor can win a server-side race and still lose the call after network transit, buffering, audio conversion, or a broken interruption.

If you need a packaged platform instead of a component API, start with the voice-agent platform comparison. If narration, accessibility, or general audio is the job, the broader text-to-speech guide uses a different decision order.

How These APIs Were Picked

This is a verified comparison, not a claim that nine APIs were exercised in one controlled lab. Every model name, public price, tier, limit, transport, and published latency number came from a live first-party page checked on October 21, 2026. The ranking then applies a voice-agent production lens to those facts.

That distinction matters because the published latency numbers do not measure the same thing. Inworld reports a server-side P90 time to first audio byte that excludes the network. ElevenLabs publishes approximate model latency that excludes application and network delay. Rime publishes P50 and P90 time to first audio at stated concurrency, then separately estimates regional network round trip. Deepgram says “as low as” 80 ms. Hume distinguishes roughly 100 ms of model latency from roughly 200 ms until first audio in instant mode. Mistral describes about 90 ms time to first audio without a percentile or network boundary on its model page; treat it as a vendor claim to validate on your route.

Those numbers are useful inside each vendor's system. They are not a clean cross-vendor benchmark.

Five criteria decided the order:

1. First playable audio: The first-audio budget starts when the agent has text ready and ends when the caller can hear speech. Full-file generation time is a different metric.
2. Interruption state: A real call needs to know what was spoken, what was not, and what the language model should retain after the caller barges in.
3. Streaming transport: Persistent WebSockets, progressive input, cancellation, and usable audio formats can remove more delay than a small model-level advantage.
4. Production economics: The relevant bill is generated characters or tokens at expected volume, plus minimums and tier effects, not the cheapest logo on a pricing page.
5. Named walls: Language coverage, number normalization, concurrency, preview status, output encoding, commercial-use rights, and unclear benchmark definitions all change the practical winner.

Voice quality still matters, but it comes after the hard failures. A beautiful voice that starts late, says an account number ambiguously, or continues talking over the caller is not a good agent voice.

1. Deepgram Flux TTS: Best Overall for Voice Agents

Deepgram Flux TTS is the strongest overall choice because its voice-agent protocol treats interruption as state, not merely a stop button.

Deepgram's product page still publishes first audio as low as 80 ms. Its GA record dates general availability to August 12, 2026, but the current interruption contract is the implementation detail to follow: stop playback locally, then send Interrupt with a session-wide playback_offset. SpeechInterrupted returns text_spoken and text_remaining only when that offset is supplied. Without it, both text fields are omitted.

Consider an account-balance agent that begins, “Your balance is two thousand eight hundred...” before the caller interrupts with a payment question. A generic cancel command can stop the waveform, but the language model may still believe the entire balance was delivered. Flux maps the supplied playback offset back to the text boundary needed to repair that state. The business consequence is fewer repeated facts, fewer contradictory follow-ups, and less custom audio-to-text reconciliation.

Flux also carries conversation context across turns through its v2 Speak WebSocket. That is useful when delivery should stay calm after a complaint or become concise after repeated interruptions. It is a better fit for a multi-turn call than a narration model that sees each line as an isolated clip.

The main wall is scope. The current catalog is 36 English voices across seven accents. If the launch requires broad multilingual coverage, Inworld or ElevenLabs moves ahead. Flux also streams raw linear16, mulaw, or alaw audio in the default Voice Agent integration, with no container or bit rate. Those formats suit telephony, but a team expecting MP3, Opus, FLAC, AAC, or WAV from the default path must change its audio assumptions or choose an Aura model.

Best for: English customer-service, scheduling, qualification, and support calls with frequent barge-in.
Standout: Native spoken-versus-unspoken text reporting after an interruption.
Pricing: Standard Pay As You Go is $0.0450 per 1,000 characters and Growth is $0.0405 per 1,000. The launch promotion has ended.
Free trial: Yes. Pay As You Go includes a free $200 credit with no minimum and no expiration.

Deepgram pricing, every current tier

The live Deepgram pricing page has three commercial routes. Pay As You Go begins with the $200 credit and supports up to 45 TTS REST or WebSocket connections. Growth starts at $4,000 per year in prepaid credits and raises that TTS concurrency ceiling to 60. Enterprise is custom for higher volume, private deployment, data, or support requirements.

Model pricing matters too. Flux Pay As You Go costs $0.0450 per 1,000 characters and Growth costs $0.0405. Aura-2 costs $0.030 and $0.027 respectively. Aura-1 costs $0.0150 and $0.0135. The cheaper Aura choices can make sense for less interactive audio, but they do not carry Flux's conversation-native interruption behavior.

A practical Flux setup

1. 1

   Open the live transport

   Connect the agent to the v2 Speak WebSocket and keep that connection alive through the conversation. Do not turn every sentence into a fresh batch request.

2. 2

   Choose the phone-ready encoding

   Select linear16, mulaw, or alaw to match the downstream audio path. Avoid a conversion step in the hot path when the telephony provider already accepts one of those formats.

3. 3

   Stream text in speakable chunks

   Send complete clauses as the language model produces them. A clause gives the voice enough context to sound natural without waiting for an entire paragraph.

4. 4

   Treat interruption as memory

   When the caller barges in, stop local playback immediately and send Interrupt with playback_offset measured from the start of the session. Discard in-flight audio. Use the returned text_spoken in conversation state; do not treat text_remaining as delivered. Keep the offset advancing across turns.

5. 5

   Measure what the caller hears

   Log text-ready time, first byte received, first frame queued, and first frame played. Optimize the slowest boundary at p95 instead of celebrating the vendor's best-case number.

Verdict: Deepgram wins this ranking because the API helps preserve conversational truth after a barge-in. Pick something else when multilingual coverage is non-negotiable or when character price is the dominant constraint.

2. Rime Mist v3: Best for Transparent Low-Latency Benchmarking

Rime Mist v3 earns second place by publishing the most decision-useful latency breakdown in this group.

Rime's latency documentation reports Mist v3 at 37 ms P50 and 56 ms P90 time to first audio with one concurrent request. At 12 concurrent requests, those published numbers remain 37 ms and 56 ms. These are measurements on a specified single H100 machine, not a phone-call benchmark.

That is more useful than a single best-case number because it exposes the tail and a load condition. It still is not your end-to-end call latency. Rime says typical network round trip adds 25 to 50 ms from much of the continental United States when the nearest region is used, while a far-coast route can add 60 to 85 ms. Your media server, audio buffer, and phone network come after that.

Mist is the speed choice; Coda v2 is the broader, more metadata-rich option. The current model comparison lists Mist with English, French, German, and Spanish and 78 voices. Coda v2 supports ten production languages and 313 voices. Both stream over HTTP and WebSockets, but Coda offers word-level timestamps while Mist does not. If the application needs precise word boundaries to reconcile playback, that missing Mist metadata can outweigh its lower TTFA.

Best for: Teams that want percentile latency data and can operate inside four Mist languages.
Standout: Published P50 and P90 TTFA at both one and 12 concurrent requests.
Pricing: Mist v3 remains $0.03 per 1,000 characters; Coda v2 is $0.05 per 1,000.
Free trial: Yes. The current page offers $50 in credits for any model.

Rime pricing, every current tier

Pay as you go charges $0.03 per 1,000 characters for Mist and $0.05 for Coda v2, with 15 concurrent Mist generations and 10 Coda generations. Enterprise uses a custom annual commitment and adds provisioned concurrency, unlimited custom voices, SLAs and dedicated support, plus cloud, on-prem, or VPC deployment.

The pricing page also expresses TTS prices per approximate minute, but says synthesis is metered by characters at about 1,000 characters per minute. The old conflicting free-minute offers have been replaced by a $50 signup credit.

Verdict: Rime Mist is the best speed specialist in the ranking. It moves above Deepgram only when the call flow can tolerate simpler interruption metadata and the real-region bakeoff confirms its percentile advantage.

3. Inworld Realtime TTS-2 Flash: Best for Global Cost and Language Coverage

Inworld Realtime TTS-2 Flash remains a strong economic choice for a multilingual voice agent, especially on paid credit plans.

Inworld's model documentation publishes a 20 ms P90 time to first audio byte for Flash and 100 ms for the more expressive TTS-2. Those numbers are server-side and exclude network latency, so they cannot be placed directly beside Rime's total cloud guidance or Hume's typical first-audio figure. The useful conclusion is narrower: Inworld has built Flash for immediate streaming over WebSocket, and the model-side boundary is unlikely to be the largest part of a well-routed call.

Coverage is the stronger differentiator. Inworld's language-specific documentation now lists 470 languages through 623 language and locale codes for both TTS-2 models. Only 94 languages fall in its supported Tier 1 and Tier 2 groups; the remaining 376 are unevaluated, best-effort languages. Its product page advertises voice cloning from 3 seconds of audio. Reusing one identity across markets is attractive, but test every target language before promising production coverage.

The cost consequence is equally sharp. On-Demand Flash is $15 per million characters, compared with $30 per million for Rime Mist, $45 per million for Deepgram Flux, and $40 per million for ElevenLabs Flash. Mistral now undercuts that on-demand rate at $13 per million, although its preview and narrower documented language scope change the decision. At the 50-million-character scenario later in this article, Inworld On-Demand is $750 per month before any volume discount.

The tradeoff is pricing complexity. Inworld uses monthly credits, model-specific usage rates, six plan levels, and different concurrency limits. A lower rate does not automatically mean a lower invoice when a paid tier's credit commitment exceeds actual usage. Match the tier to real consumption instead of buying a discount percentage.

Best for: High-volume multilingual agents and one voice identity across many languages.
Standout: Broad multilingual coverage, WebSocket streaming, and a $15-per-million-character on-demand Flash rate.
Pricing: Start free; Flash falls from $15 per million characters on On-Demand to below $5 on Enterprise.
Free trial: Yes. On-Demand includes up to 70 minutes of TTS.

Inworld pricing, every current tier

The live Inworld pricing page uses credits to cover TTS, STT, and language-model usage:

• On-Demand: Starts free, includes up to 70 TTS minutes, charges Flash at $15 per million characters and TTS-2 at $25, with 5 concurrent requests.
• Creator: $25 per month in credits, Flash at $10 per million and TTS-2 at $20, with 10 concurrent requests.
• Builder: $100 per month in credits, Flash at $9 per million and TTS-2 at $17.50, with 50 concurrent requests.
• Developer: $300 per month in credits, Flash at $8 per million and TTS-2 at $15, with 150 concurrent requests.
• Growth: $1,500 per month in credits, Flash at $7 per million and TTS-2 at $12.50, with 500 concurrent requests.
• Enterprise: Custom, with TTS-2 as low as $5 per million, Flash below $5 per million, and custom concurrency.

Inworld's own calculator assumes roughly 1,000 characters per generated minute. Use that convention for a first budget, then replace it with the measured character count from actual prompts. A terse collections agent and a verbose tutoring agent will not consume the same characters per minute.

Verdict: Inworld is the first pick when the launch is global or character spend is a board-level line item. Deepgram stays ahead for English calls where native barge-in state saves more engineering and conversational risk than the price delta.

4. ElevenLabs Flash v2.5: Best Mature Multilingual Ecosystem

ElevenLabs Flash v2.5 is the safest mainstream choice when voice selection, multilingual operations, and a familiar API ecosystem matter together.

ElevenLabs' model documentation publishes roughly 75 ms of latency for Flash v2.5, excluding application and network delay. The model supports 32 languages and a 40,000-character request limit. For a voice agent, the practical appeal is not one isolated number. It is the combination of low-latency positioning, broad language support, established voice operations, and a public pay-as-you-go path.

The important wall hides in text normalization. Flash disables number normalization by default to preserve latency. Phone numbers, dates, and currencies may be spoken in a way the caller does not expect. Enterprise customers can turn normalization on for v2.5. Everyone else should normalize those values in the language model before passing text into TTS.

That is not an edge case. A support agent regularly speaks order IDs, dates, amounts, addresses, confirmation codes, and phone numbers. The fastest voice is useless if a compact date string comes out as one giant number rather than a date. Put normalization into the prompt and test suite, not into a last-minute production patch.

Best for: Multilingual products that value an established voice ecosystem and simple API procurement.
Standout: Roughly 75 ms published model latency across 32 languages.
Pricing: Flash and Turbo now cost $0.04 per 1,000 characters.
Free trial: Yes. Free / Pay As You Go includes 20,000 Flash or Turbo characters.

ElevenLabs pricing, every current tier

The live ElevenLabs API pricing page lists:

• Free / Pay As You Go: $0 with 20,000 Flash or Turbo characters.
• Starter: $6 per month with 150,000 characters.
• Creator: $22 per month, with the first month at $11, and 550,000 characters.
• Pro: $99 per month with 2,475,000 characters.
• Scale: $299 per month with 7,475,000 characters.
• Business: $990 per month with 24,750,000 characters.
• Enterprise: Custom.

The Flash rate is now $0.04 per 1,000 characters across the displayed API tiers. That makes the plan decision about included usage, account features, and operational limits rather than a lower displayed per-character rate.

Verdict: ElevenLabs is a defensible default, not the automatic number one. Pick it when voice operations and multilingual familiarity outweigh the extra character cost, then make normalization a release blocker.

5. Mistral Voxtral TTS 3: Best Low-Cost Preview Candidate

Mistral Voxtral TTS 3, a text-to-speech model, enters at fifth for builders willing to validate a preview before committing production calls.

Since October 20, 2026, Voxtral TTS 3 has been in public preview. Its model page states streaming with about 90 ms time to first audio, nine languages, and $13 per million characters. Use the model page's language scope when planning a launch; the broader release announcement is not the specification for this entry.

The practical wall is input streaming. The streaming guide requires complete text in each request and streams the audio back as server-sent events. That can suit an agent that sends finished clauses, but it is not a persistent session accepting text continuously as the language model writes. The TTS overview lists a separate realtime model; its capabilities should not be silently credited to voxtral-mini-tts-3-0.

For a scheduling agent speaking short confirmations in a supported language, the standard character rate makes Voxtral worth a pilot. It stays behind the top four because preview status, the complete-text request boundary, and unproven interruption reconciliation matter more than saving a few milliseconds on a model card. Its roughly 90 ms claim also lacks the percentile and measurement boundary needed to declare it faster than another vendor's API.

Best for: Cost-sensitive pilots that can send finished text chunks and tolerate preview changes.
Standout: Streamed audio with a low standard character rate.
Pricing: $13 per million characters, before any separately qualified caching discount.
Free trial: Not confirmed in the checked model and pricing documentation.

Verdict: Pilot Voxtral when its documented scope fits. Promote it only after its first-playback latency, pronunciation, and interruption handling pass the same call tests as the established candidates.

6. Cartesia Sonic-3.6: Best for Persistent WebSocket Contexts

Cartesia Sonic-3.6 is the best fit for developers who want explicit WebSocket context control, cancellation, and timestamp events.

Cartesia's live pricing table now names Sonic-3.6. Its WebSocket API carries a context identifier, and the same connection supports context cancellation plus word and phoneme timestamp responses. That is useful when an agent streams several clauses, needs to stop one utterance cleanly, and wants fine-grained playback metadata.

The honest latency finding is an omission: the current Sonic-3.6 model documentation and pricing page do not publish a specific TTFA number. Older Cartesia latency figures attached to earlier Sonic versions should not be silently recycled onto 3.6. That keeps Cartesia below vendors with current, attributable latency boundaries, even though the transport design is compelling.

Pricing is credit-based and the current page displays monthly subscriptions. The minutes are approximate, which is fine for an initial budget but should be replaced by measured consumption before a contract.

Best for: Developers who need context IDs, cancellation, and timestamp-rich WebSocket output.
Standout: Word and phoneme timestamps plus explicit context cancellation.
Pricing: Free starts at about 27 minutes; current paid monthly rates begin at $5 per month.
Free trial: Yes. The Free tier includes 20,000 monthly credits, shown as about 27 TTS minutes.

Cartesia pricing, every current tier

• Free: $0 per month, about 27 TTS minutes, and 2 concurrent requests.
• Pro: $5 per month on monthly billing, about 133 minutes, and 3 concurrent requests.
• Startup: $49 per month on monthly billing, about 1,667 minutes, and 5 concurrent requests.
• Scale: $299 per month on monthly billing, about 10,667 minutes, and 15 concurrent requests.
• Enterprise: Custom credits, agent usage, volume pricing, and concurrency.

Cartesia's model naming and pricing are current enough to buy against. Its latency marketing is not specific enough to compare Sonic-3.6 confidently. Make the vendor earn that place with a real-region bakeoff.

Verdict: Cartesia is a strong transport-first shortlist candidate. It rises only after its current model beats the leaders on your measured p95 path, because an old Sonic number is not a Sonic-3.6 fact.

7. Gradium: Best for a Unified EU and US Speech Stack

Gradium is the pragmatic choice when one provider for TTS and STT, automatic regional routing, and predictable bundles matter together.

Gradium's API reference supports both REST and WebSocket speech endpoints. Requests go to one API hostname and are automatically routed to the nearest EU or US cluster. The default hostname carries no residency commitment. For guaranteed regional processing, enroll in data residency and use the EU- or US-pinned endpoint; the API reports the residency in each response. Nearest-region routing alone is not a data-residency guarantee.

The current TTS product page now cites 49 ms P50 TTFA from Speko and 50 ms perceived P50 from Coval. It also displays about 200 ms in a real-time API panel, so these are attributed benchmark and integration figures, not one universal guarantee. Measure the exact transport and region you intend to deploy.

The pricing model is monthly credits rather than a simple character rate. One TTS character consumes one credit, and Gradium maps about 45,000 characters to one hour of speech. The bundle can be attractive if the same account uses transcription or translation, but the comparison becomes less clean when TTS is the only workload.

The sharpest entry-tier wall is commercial rights. Free includes API access and about one hour of TTS, but commercial use is not allowed. The $13 XS plan is the first public commercial tier.

Best for: Products that want one regional vendor for synthesis and transcription.
Standout: One API hostname with automatic EU and US cluster routing.
Pricing: Free for noncommercial use; commercial plans start at $13 per month.
Free trial: Yes. Free includes 45,000 credits, about one TTS hour, and 2 TTS concurrency.

Gradium pricing, every current tier

The current Gradium pricing page lists:

• Free: $0 per month, 45,000 credits, about 1 TTS hour, 2 concurrency, API access, and no commercial use.
• XS: $13 per month, 225,000 credits, about 5 TTS hours, 5 concurrency, and commercial use.
• S: $43 per month, 900,000 credits, about 20 TTS hours, 5 concurrency, and commercial use.
• M: $340 per month, 9 million credits, about 200 TTS hours, 10 concurrency, and commercial use.
• L: $1,615 per month, 45 million credits, about 1,000 TTS hours, 15 concurrency, and commercial use.
• Enterprise: Custom with unlimited included credits and TTS plus custom concurrency.

Additional 100,000-credit packs cost $6.90 on XS, $5.00 on S, $4.00 on M, and $3.80 on L. That declining overage ladder rewards committing to a larger bundle, but it also means the effective rate depends on utilization.

Verdict: Gradium is a strong stack-consolidation shortlist candidate with current low-latency benchmark claims. Confirm the actual call path and enroll in residency if regional processing is a requirement.

8. OpenAI GPT-4o Mini TTS: Existing Integrations Only

OpenAI GPT-4o Mini TTS is now an existing-integration bridge, not a fresh production choice. OpenAI's deprecation notice schedules shutdown for January 6, 2027.

The model accepts text and produces audio through the official speech endpoint. The API can stream audio or server-sent events, offers 13 named built-in voices plus custom voice IDs, and supports MP3, Opus, AAC, FLAC, WAV, and PCM output. Playback speed can range from 0.25 to 4.0.

The missing fact is a current official TTFA figure. OpenAI's model page does not publish one, so GPT-4o Mini TTS cannot rank as a low-latency leader on documentation alone. Its streaming endpoint still matters to existing deployments, but provider familiarity no longer justifies starting a new integration on a deprecated model.

Pricing also resists a clean character comparison. OpenAI bills $0.60 per 1 million input text tokens and $12 per 1 million output audio tokens. That can be reasonable, but a buyer must forecast from observed token use rather than multiplying a public character rate. The model page lists a 2,000-input-token maximum.

Best for: Existing Speech API customers completing a migration before shutdown.
Standout: Flexible audio formats, streaming modes, voice instructions, and provider consolidation.
Pricing: $0.60 per million input text tokens plus $12 per million output audio tokens.
Free trial: No. The model's Free usage tier is not supported.

OpenAI usage tiers, every current limit

The official GPT-4o Mini TTS model page lists no Free access. Build allows 2,000 requests per minute and 150,000 tokens per minute. Launch allows 10,000 and 2 million. Grow allows 10,000 and 8 million.

Those are usage limits, not monthly subscriptions. The immediate production question is how to migrate without interrupting existing users.

Verdict: Migrate existing users before shutdown. The speech guide recommends gpt-realtime-2.1-mini through the Realtime API; this requires a workflow change, not a model-name swap on the Speech endpoint.

9. Hume Octave 2: Existing Deployments Should Plan Their Exit

Hume Octave 2 should not be chosen for a new production voice agent: Hume's TTS and EVI APIs close on November 13, 2026 at 12:01 a.m. EST. Existing APIs remain supported until then, so the useful job is migration planning.

Hume's Octave documentation still marks Octave 2 as a preview and publishes model latency around 100 ms excluding network transit. Instant mode typically makes first audio ready in roughly 200 ms depending on load and input complexity. Those are different boundaries, and neither changes the shutdown decision.

The preview supports English, Japanese, Korean, Spanish, French, Portuguese, Italian, German, Russian, Hindi, and Arabic. It can clone a voice from as little as 15 seconds of audio, with HTTP output streaming, bidirectional WebSocket streaming, and non-streaming responses.

Hume's semantic and emotional delivery can remain part of an existing application's migration acceptance tests. A coaching or sensitive customer-care flow should preserve the delivery qualities that matter while moving to a supported provider.

Best for: Existing customers validating a replacement before shutdown.
Standout: Expressive delivery to compare against replacement voices.
Pricing: A current public TTS rate could not be verified; use the existing account's billing terms during migration.
Free trial: Not reverified; the shutdown makes it unsuitable as a new production trial.

Hume pricing and availability

The former pricing page now redirects to Hume's corporate homepage, so the old subscription and overage ladder is no longer a reliable buying reference. Check the current shutdown notice and complete the migration before access ends.

Verdict: Retain Hume only long enough to validate and complete an existing deployment's migration. It no longer earns an expression-led recommendation for a new build.

What 50,000 Generated Minutes Cost

The cleanest budget comparison uses a common character assumption and public on-demand rates. Inworld's pricing calculator assumes about 1,000 characters per minute, so 50,000 generated minutes becomes 50 million characters. This is a scenario, not a promised invoice.

At that volume, before signup credits or negotiated discounts:

The August illustration below retains the old ElevenLabs amount. The table above is the current calculation: ElevenLabs is $2,000, its gap to Inworld On-Demand is $1,250, and Mistral adds a $650 option.

Mistral is the lowest standard on-demand rate in this table, but paid plans can reverse the result. At Inworld's Developer rate of $8 per million characters, the same workload costs $400: $300 in monthly credits plus $100 of overage. That is why a low preview rate alone does not displace Inworld for a production buyer.

Do not turn that gap into an automatic recommendation. A cheaper voice that increases repeat calls, misreads account numbers, or forces weeks of custom interruption work can cost more than it saves. Conversely, paying a familiar vendor premium without measuring a business benefit is just inertia.

The scenario has three deliberate simplifications. First, characters per minute vary with language, speech rate, and punctuation. Second, paid credit plans can create minimum spend or unlock lower usage rates. Third, enterprise discounts are private. Put one week of real agent transcripts through a character and token counter before signing an annual commitment.

Who Should Pick What

Pick Deepgram Flux TTS when English coverage is enough and callers regularly interrupt. Native spoken-state reporting is the decision-flipping capability.

Pick Rime Mist v3 when the fastest transparent percentile profile matters, the four supported languages cover the launch, and the application can work without Mist word timestamps. Put media servers near users, because the network can erase the model's advantage.

Pick Inworld TTS-2 Flash when the same agent must speak across many markets or when generated-character spend is the largest TTS constraint. Its paid credit tiers can beat the standard on-demand quotes in the 50,000-minute scenario; qualify each target language against the documented support tiers.

Pick ElevenLabs Flash v2.5 when the organization wants a mature multilingual voice operation and a simple path from free usage to larger plans. Normalize every number, date, currency, code, and address before synthesis.

Pick Mistral Voxtral TTS 3 for a price-sensitive pilot within its nine documented languages when complete-text requests fit. Validate first playback and interruption handling before promoting the preview to production.

Pick Cartesia Sonic-3.6 when context cancellation and word or phoneme timing matter to the playback architecture. Require a fresh TTFA result from the real deployment, because the current model pages do not provide one to inherit.

Pick Gradium when one TTS and STT bill, automatic EU or US routing, and a regional speech stack remove operational work. Start on XS for any commercial prototype because Free is noncommercial.

Migrate OpenAI GPT-4o Mini TTS deployments before January 6, 2027. Existing users should evaluate the recommended Realtime API workflow instead of extending their dependency on the deprecated Speech model.

Migrate Hume Octave 2 deployments before November 13, 2026. Its expressive delivery is a migration test case, not a reason to start a new integration.

One question flips most close calls: What happens when the caller interrupts halfway through a consequential fact? If the answer is “the client guesses what played,” fix that before debating voice taste.

The Ones to Avoid for Latency-First Voice Agents

Pika Speech: Fast Full Clips, Wrong Live Transport Today

Pika Speech is an impressive narration and voice-cloning system, but its current API is the wrong transport for a latency-first live voice agent.

Pika's release post reports a 0.02 real-time factor in locally run three-minute tests. Under its long-form benchmark, a minute of speech takes about 1.2 seconds to generate. The model produces 48 kHz audio, can clone from 5 seconds of reference audio, and accepts up to 5 minutes of speech per request.

That is excellent full-clip throughput. It is not evidence that a caller hears the first syllable quickly.

The live Pika Speech API documentation explicitly lists real-time streaming synthesis under “Not for.” A POST returns a queued job, and the client polls until the job is complete. The vendor's benchmark table records Pika Speech at $0.01 per generated minute as of August 14, 2026, which is a historical benchmark price rather than a verified current checkout quote. The asynchronous contract still disqualifies it from this live-turn ranking today.

Best for: Narration, voiceover, and fast full-clip voice generation.
Standout: About 1.2 seconds to generate one minute under Pika's long-form benchmark.
Pricing: Historical benchmark: $0.01 per generated minute as of August 14, 2026. Current checkout pricing was not confirmed.
Free trial: Not confirmed on the checked model and release pages.

Verdict: Use Pika for queued speech jobs today. Re-evaluate it for live agents only when the public API exposes a real streaming path with a measured first-audio boundary.

Avoid “free unlimited” as a production requirement

Several providers offer real free credits or free monthly allowances. None of the checked pages promises unlimited, production-grade, low-latency TTS at no cost. Free plans also carry walls such as noncommercial use, low concurrency, unsupported access, or a small character balance.

Budget a paid path before the prototype works. Otherwise success itself becomes the incident.

Avoid the quality model when speed is the hard constraint

Most vendors now separate a fast conversational model from a more expressive or long-form model. Choose Flash over a slower ElevenLabs option for latency-first calls, Mist over Coda when raw TTFA is decisive, and Inworld Flash over TTS-2 when cost and speed matter more than maximum expression. Do not pay the latency tax for quality a phone line or use case cannot exploit.

The Monday Move

Next Monday, run a five-scenario bakeoff with Deepgram Flux, Rime Mist, Inworld Flash, and ElevenLabs Flash. Add Voxtral TTS 3 if its nine-language scope and complete-text request boundary fit the product. Use the same language model output, media server, region, phone carrier, audio encoding, and playback buffer for every candidate.

The five scenarios should expose different failures:

1. A short greeting that reveals first-audio delay.
2. A long answer interrupted halfway through a sentence.
3. A turn containing a person's name, an August date, a currency amount, phone number, and confirmation code.
4. A streamed language-model response with uneven chunk sizes.
5. A concurrency burst that reaches the expected launch peak.

Record text-ready time, first byte received, first frame queued, first frame played, p50 TTFA, p95 TTFA, interruption correctness, and the exact text the caller heard. Score pronunciation and voice preference only after every candidate clears the hard latency and barge-in threshold.

Then attach the real generated character count to the pricing tier the deployment would actually buy. The Monday decision is not “which demo sounds best?” It is “which survivor produces correct, interruptible calls at the lowest complete monthly cost?”

If English coverage is enough, Deepgram's interruption state makes it the default. If the product is multilingual, start with Inworld and ElevenLabs. If raw first audio is the gating metric, keep Rime in the final pair. One day of disciplined measurement is more valuable than another week comparing vendor adjectives.

Frequently Asked Questions

What is the best TTS API?

Deepgram Flux TTS is the best overall API in this comparison for an English voice agent because it combines low published first audio with native interruption state. Inworld is the better answer when multilingual reach or paid-plan character economics matter more, while Rime is the speed specialist.

Is there a free low-latency text-to-speech API?

Yes. Deepgram offers a $200 Pay As You Go credit, Rime offers $50 in credits, Inworld includes up to 70 TTS minutes on On-Demand, ElevenLabs includes 20,000 Flash characters, Cartesia includes about 27 minutes, and Gradium about one hour for noncommercial use. These are prototype allowances, not free unlimited production plans.

What is the cheapest TTS API?

At 50 million characters, Voxtral TTS 3 is cheapest among the five standard on-demand rates compared here at $650. Inworld Flash on the paid Developer plan is lower at $400 for the same usage. Compare preview risk, language coverage, commitments, and real transcript counts before choosing.

What is the best local TTS model?

A local or self-hosted model is a different procurement decision from a managed API. Deepgram Flux and Rime both document private deployment routes, but the winner depends on your hardware, concurrency, model access terms, and ability to operate the serving stack. Benchmark first audio and real-time factor on the exact hardware you intend to own.

Does a low-latency TTS API work on Android?

Yes. Keep the vendor API key on your backend, stream generated audio from that backend to the Android client, and measure first playback on the device. Embedding a secret in the app is unsafe, and mobile buffering or network conditions can dominate the provider's model-level latency.

Get the AI tools map for business owners to compare the rest of your agent stack without rebuilding this research from scratch.