Best Text-to-Speech APIs for Real-Time Voice Agents in 2026

A voice agent can have the best language model and the best transcription and still lose the caller if the voice is slow, flat, or mispronounces an order number. Text-to-speech decides two things the caller notices immediately: how long the line stays silent after they stop talking, and whether the reply sounds like a person who cares. Prices for the same million characters now run from $4 to $80, and the quality ranking has moved twice since August.

The short answer: on the independent Artificial Analysis Speech Arena (fetched Nov 11, 2026), ElevenLabs Eleven v4 and v4 Turbo hold ranks one and two, and Cartesia Sonic 3.6 is fourth. For the lowest bill at mainstream quality, Deepgram Aura-2 and Rime Mist v3 list at about $30 per 1M characters. Two incumbents are on their way out: OpenAI removes tts-1, tts-1-hd and gpt-4o-mini-tts on Jan 6, 2027, and Hume ends TTS API access on Nov 13, 2026. If you are choosing the whole stack, read the voice agent platforms comparison, pair this guide with the speech-to-text APIs guide, and if you want a worked build, see how to build an AI phone calling agent.

A Deepgram-produced comparison of its own TTS against ElevenLabs Flash. It is vendor marketing (it claims Deepgram is about 40% cheaper than ElevenLabs Flash), so use it for the use case framing, not as evidence.

Quick Picks

These are the shortlists we would start from. Each line names the criterion, the pick, and the evidence behind it. All prices and scores are as of Nov 11, 2026.

Master Comparison Table

One row per product. Prices are per 1M characters because that is the unit most of the market quotes; where a vendor bills per minute, per token or per credit, the cell shows the conversion and says so. Where a vendor did not publish a number on the pages we fetched, the cell says that instead of guessing. The last column is the Artificial Analysis Speech Arena Elo and rank among 104 models.

Reading notes: Artificial Analysis (AA) states that its prices reflect 1M characters at default settings and converts token-priced models for comparison, so Gemini, OpenAI and Azure cells marked “AA” are AA's conversion, not a vendor list price. Latency figures in this table are what vendors say about themselves. The latency section adds an independent measurement and the two do not agree.

ElevenLabs (v4, v4 Turbo, v3, Flash)

ElevenLabs sells the widest menu of voices and models in this guide. On Sep 28, 2026 it launched Eleven v4, tuned for produced content, and Eleven v4 Turbo, a lower-latency variant for agents. Both support 90+ languages and inline audio tags such as [laughs], and both support Professional Voice Clones, which v3 did not. ElevenLabs says Turbo has a median inference latency of about 100 ms and a median time to first speech of about 150 ms (vendor-reported), and that it trades slightly lower audio quality for that speed. It also says Turbo was optimized together with its own ElevenAgents platform, which matters if you plan to run it elsewhere.

Pricing (as of Nov 11, 2026). The API pricing page bills per 1,000 characters in USD: Eleven v4 at $0.08 ($80 per 1M), Eleven v4 Turbo at $0.04 ($40 per 1M), Eleven v3 at $0.08, and the Flash and Turbo tier at $0.04. The launch discount of $22 and $11 per 1M characters ran for two weeks after Sep 28 and is not what the page lists now. Subscription plans run Starter $6, Creator $22, Pro $99, Scale $299 and Business $990 per month, and Enterprise adds SLAs and SOC 2. Its Realtime v1 speech-to-speech models list 620 ms (Mini) and 570 ms (Max) time to first audio (vendor-reported) at $0.08 per minute plus LLM usage.

Model limits worth knowing. Per ElevenLabs' models documentation, v4 accepts 10,000 characters per request, v3 accepts 5,000, Multilingual v2 supports 29 languages at 10,000 characters, and Flash v2.5 supports 32 languages at 40,000 characters at a lower per-character price. The docs and the pricing page disagree on some latency labels: Flash is about 100 ms in the docs and about 75 ms on the pricing page, and v3 Conversational is about 165 ms in the docs and about 280 ms on the pricing page. We report both rather than pick one.

# ElevenLabs REST call (endpoint and fields from ElevenLabs' API reference)
curl --request POST \
  --url "https://api.elevenlabs.io/v1/text-to-speech/JBFqnCBsd6RMkjVDRZzb?output_format=mp3_44100_128" \
  --header "xi-api-key: $ELEVENLABS_API_KEY" \
  --header "Content-Type: application/json" \
  --data '{"text": "The first move is what sets everything in motion.", "model_id": "eleven_v4"}' \
  --output reply.mp3
# JBFqnCBsd6RMkjVDRZzb is the voice ID in ElevenLabs' own example; use your own voice ID.
# model_id eleven_v4 comes from ElevenLabs' v4 blog post. Check the models docs for the Turbo id.
# For phone audio, output_format accepts ulaw_8000; for live agents use the streaming endpoint.

Strengths. Top two spots on the independent arena (Elo 1338 and 1337), the broadest language list (90+), mature voice cloning with both instant (about 10 seconds of audio) and professional tiers, and the most expressive control through tags and natural-language direction. Turbo is priced at half of v4 and sits one Elo point below it, which is the strongest value signal in this guide.

Weaknesses. Turbo's independent time to first audio is 189 ms median and 239 ms at P95 in the Cekura data we read, which is competitive but behind the specialists that publish 50 to 60 ms medians. Cartesia's comparison page claims ElevenLabs latency varies widely from call to call and that its realtime models support fewer languages (32); both claims are Cartesia's and we did not test them. One third-party guide also reports that the ElevenLabs WebSocket API waits for about 120 characters by default before generating the first chunk, which affects streamed LLM output; we have not verified that setting. Best for: agents where tone, accent authenticity or 90+ languages matter more than shaving the last 50 ms, and teams that want a cloned brand voice.

Cartesia Sonic 3.6

Cartesia built Sonic on state space models rather than the transformer designs most rivals use, and positions it as the speed-first option for voice agents. Sonic 3.6 became generally available on Aug 27, 2026, and its docs list 44 languages for the dated snapshot sonic-3.6-2026-08-27. Cartesia says it replies in under 90 ms (vendor-reported), is deployable through Baseten, Amazon SageMaker or Together AI so audio stays in your own cloud, and carries a 99.9% uptime SLA, with SOC 2, PCI, HIPAA and GDPR coverage. Its launch post also cites 93% blind preference across fifteen locales, a vendor test.

Pricing (as of Nov 11, 2026). Cartesia bills in credits at one credit per character. Free is $0 with 20K credits a month; Pro is $5 for 100K credits; Startup is $49 for 1.25M; Scale is $299 for 8M. Overage is $65, $45 and $38 per 1M credits on Pro, Startup and Scale. That makes the effective rate about $50, $39 and $37 per 1M characters, and Artificial Analysis lists $49. Cartesia says one minute of audio uses roughly 750 to 800 credits. Its voice agent product is $0.03 per minute and telephony is $0.014 per minute. Pro Voice Cloning training is free, cloned speech costs the same one credit per character, and localizing a voice to a new accent costs 225 credits.

# Cartesia request shape (from Cartesia's Hume migration guide)
POST https://api.cartesia.ai/tts/bytes
  X-API-Key: <your key>
  Cartesia-Version: 2026-08-14
  body: model_id = "sonic-3.6"      # moves with new stable snapshots
        # or pin a dated snapshot: "sonic-3.6-2026-08-27" (never changes)
        # plus transcript, voice and output_format fields (see Cartesia docs)
# /tts/bytes takes one transcript per request. For multi-turn calls, use the
# WebSocket API and its context feature to keep prosody continuous.

Strengths. Fourth on the arena (Elo 1296, up from the 1282 Cartesia cited in August), instant cloning from about 10 seconds and professional clones from 30 minutes, and dated snapshots so a model update never changes your production voice without notice. Older professional voice clones carry over to 3.6 per the docs. Cartesia also publishes a migration guide for teams moving off Hume.

Weaknesses. The independent measurement does not match the headline: Cekura data read via voiceagent-tts.com shows Sonic 3.6 at 286 ms median and 374 ms at P95 time to first audio, versus the under 90 ms claim. The two numbers may count different things, so run your own test. ElevenLabs v4 and v4 Turbo outscore it on the arena, and the plan ladder puts concurrency caps on lower tiers (a third-party guide lists 2, 3, 5 and 15 concurrent requests from Free to Scale; confirm on Cartesia's page). Cartesia's ElevenLabs comparison page reports 92% blind preference over Eleven v3 in US English, which is a self-run test against an older ElevenLabs model. Best for: latency-sensitive phone agents, regulated teams that need a VPC or on-prem route, and teams that want one vendor for STT (Ink) and TTS.

A creator demo of Cartesia's earlier Sonic 3 model. Sonic 3.6 is newer, and the title is the creator's opinion, so use the video for the sound of the voices only.

Deepgram (Aura-2, Flux TTS)

Deepgram is best known for speech-to-text, and its TTS line is built around keeping the whole voice stack on one vendor. Aura-2 is its mainstream model; the newer Flux TTS sits at a higher price. Aura-2 covers English in five regional variants (American, British, Australian, Irish, Filipino), Spanish in four (Mexican, Peninsular, Colombian, Latin American), and German, French, Dutch, Italian and Japanese, according to its models documentation. The models page documents no voice cloning.

Pricing (as of Nov 11, 2026). Per 1,000 characters on pay-as-you-go and Growth plans: Aura-2 is $0.030 and $0.027 ($30 and $27 per 1M), Aura-1 is $0.015 and $0.0135, and Flux TTS is $0.045 on both ($45 per 1M), with a promo of a $1 credit per $1 spent up to $500 through Dec 31, 2026. New accounts get a $200 credit. Its Voice Agent API is $0.075 per minute on Standard and $0.065 if you bring your own TTS. Deepgram lists SOC 2, a HIPAA BAA on enterprise plans, an EU endpoint at api.eu.deepgram.com, and self-hosted containers on NVIDIA GPUs for enterprise.

# Deepgram Aura-2 REST call (model name from Deepgram's TTS models docs)
curl --request POST \
  --url "https://api.deepgram.com/v1/speak?model=aura-2-thalia-en" \
  --header "Authorization: Token $DEEPGRAM_API_KEY" \
  --header "Content-Type: application/json" \
  --data '{"text": "Your order ships tomorrow."}' \
  --output reply.mp3
# EU data residency: use api.eu.deepgram.com instead of api.deepgram.com

Strengths. One of the two cheapest mainstream voices ($30 per 1M, $27 on Growth), simple REST, residency options, and a single bill if you already use Deepgram for STT (see our STT comparison). In the Cekura data Flux TTS shows 198 ms median and 267 ms at P95 time to first audio, which is near ElevenLabs Turbo.

Weaknesses. Quality is the trade: Flux TTS is 71st on the arena (Elo 1052) at $45, and the same data shows Aura-2 at 263 ms median but 525 ms at P95. Deepgram does not publish a single latency figure on the pages we read, no cloning is documented, and its language list is the shortest here at seven. A third-party guide lists concurrency of 45 streams on pay-as-you-go. Best for: cost-sensitive, English-first agents and teams that value one vendor for STT, TTS and orchestration.

OpenAI TTS (deprecated)

OpenAI's speech endpoint was the default for many prototypes: 11 built-in voices, an instructions field that steers tone, streaming output, and language support that follows Whisper's list of about 57 languages with voices optimized for English. It is also the clearest migration risk in this guide. OpenAI's deprecations page (fetched Nov 12, 2026) lists tts-1, tts-1-hd and gpt-4o-mini-tts, including both gpt-4o-mini-tts snapshots, for removal on Jan 6, 2027. The stated replacement is gpt-realtime-2.1-mini via the Realtime API, and the TTS guide warns that it is not a supported model on /v1/audio/speech, so this is a workflow change, not a model-name swap.

// OpenAI Speech API (deprecated: shuts down January 6, 2027)
const mp3 = await openai.audio.speech.create({
  model: "gpt-4o-mini-tts",
  voice: "coral",
  input: "Today is a wonderful day to build something people love!",
  instructions: "Speak in a cheerful and positive tone.",
});
// OpenAI's replacement path is gpt-realtime-2.1-mini through the Realtime API.
// It is not supported on /v1/audio/speech, so this is a workflow change.

Pricing (as of Nov 11, 2026). tts-1 is $15 per 1M characters and tts-1-hd is $30. gpt-4o-mini-tts bills $0.60 per 1M text tokens in and $12 per 1M audio tokens out, so it has no per-character list price. The Realtime replacement bills audio tokens: gpt-realtime-2.1 is $32 in and $64 out per 1M audio tokens, and gpt-realtime-2.1-mini is $10 and $20. AA converts GPT-Realtime-2 to $191.56 per 1M characters, though that figure covers a speech model that also reasons, not TTS alone.

Strengths. Simple API and wide language coverage, approved custom voices from a consent recording and audio sample, and a single vendor if the LLM is already OpenAI's. Weaknesses. Arena Elo is 1106 for tts-1-hd and 1085 for tts-1, far below the leaders, the models are scheduled for removal, and OpenAI's usage policy requires telling end users the voice is AI-generated. We found no published latency figure for these models. Best for: existing integrations that need a bridge until the migration deadline. Do not start new builds on them.

Google Cloud (Chirp 3, Gemini TTS)

Google sells two overlapping families. Cloud Text-to-Speech has the classic tiers plus Chirp 3: HD, which its documentation lists with 30 voices (for example Kore and Puck) in about 50 locales, pace control in every locale, pause control in most, and custom pronunciation in many. Voice names follow the pattern en-US-Chirp3-HD-Kore. Separately, the Gemini TTS models are token-priced and currently in Preview.

Pricing (as of Nov 11, 2026). Chirp 3: HD is $30 per 1M characters with the first 1M free, and instant custom voice is $60. WaveNet and Standard are $4 per 1M with 4M free, Neural2 and Polyglot are $16 with 1M free, and Studio is $160. Gemini 3.8 Flash TTS (Preview) is $0.50 per 1M text tokens in and $9.00 per 1M audio tokens out through Dec 31, 2026, then $1.00 and $18.00 from Jan 1, 2027. Gemini 3.8 Flash-Lite TTS is $0.50 and $6.00, rising to $1.00 and $12.00. Google counts 25 audio tokens per second, so one minute of speech is about 1,500 tokens: our arithmetic puts Flash TTS near $0.0135 per minute of audio today and double that in 2027, before text tokens. AA lists $16.49 and $11.03 per 1M characters.

Strengths. Gemini 3.8 Flash TTS is 6th on the arena (Elo 1283) and Flash-Lite is 11th (1259) at a fraction of ElevenLabs' price, which makes them the best quality-per-dollar option in the table. Chirp 3: HD brings the broadest locale list among hyperscalers and Google Cloud's IAM and regional controls. Weaknesses. Chirp 3: HD itself scores 1061 (65th), Gemini TTS is Preview with a price rise scheduled for Jan 1, 2027, and Google does not publish a time-to-first-audio figure on the pages we read. We found no cloning detail for Gemini TTS in them, so verify before you plan on it. Best for: Google Cloud shops, multi-locale deployments, and budget-conscious teams willing to accept a Preview label.

Azure Speech

Azure's speech service groups TTS into Neural, Neural HD, Neural HD Flash, MAI-Voice, Professional Voice and Personal Voice, the last of which is limited access. The free F0 tier includes 0.5M neural characters a month, and commitment tiers (the page mentions 80M characters) discount high volume. Microsoft's language page lists a very large locale table, and we did not count it.

Pricing. Azure's pricing page rendered dollar values as placeholders when we fetched it, so we cannot give you a Microsoft-confirmed per-1M price. Artificial Analysis lists Azure Neural at $15 and Azure HD 2.5 at $22 per 1M characters. Strengths. Enterprise procurement, regional availability, and custom voice options; Azure HD 2.5 is 34th on the arena (1145). Weaknesses. Azure Neural scores 1032 (80th), the pricing is hard to read from the public page, and Personal Voice requires approval. We found no independent latency figures for Azure voices in the data we read. Best for: organizations already standardized on Azure and its compliance posture.

Amazon Polly

Polly is the cheapest and oldest service here. Its engines list at $4 per 1M characters for Standard, $16 for Neural, $30 for Generative and $100 for Long-Form. The free tier gives 5M Standard characters a month, 1M Neural characters for the first 12 months, 500K Long-Form, and 100K Generative for 12 months. Its language page lists 40 language and locale rows, including Welsh, Icelandic, Catalan and Cantonese, which few competitors cover.

Strengths. Lowest price, deep AWS integration, and breadth of languages. Artificial Analysis lists Polly Standard as the fastest throughput of all models at 936.9 characters per second via Bedrock, though that is generation speed, not time to first audio. Weaknesses. Quality: Polly Generative is 62nd (1063), Neural is 98th (892) and Standard is 104th (830), last of 104. Expressive control and cloning are not documented in the pages we read. Best for: IVR, notifications and low-stakes prompts where intelligibility matters more than personality, or as a fallback voice when your primary vendor is down.

Rime (Mist v3, Coda v2)

Rime sells speed and phone-call realism with two models. Mist v3 covers English, French, German and Spanish with 78 voices and inline pronunciation control in English. Coda v2 covers 10 languages (English, Spanish, Portuguese, French, German, Italian, Dutch, Japanese, Arabic, Hindi) with 313 voices and word-level timestamps. Audio streams over HTTP and WebSockets. Rime reports a Mist v3 time to first audio of 37 ms P50 and 56 ms P90 at one concurrent request (vendor-reported), about 70 ms when self-hosted, and under 100 ms P50 for Coda v2.

Pricing (as of Nov 11, 2026). Rime bills per minute: Free is $0 with $50 of usage a month, and Flex pricing is $0.030 per minute for Mist v3 and $0.050 per minute for Coda v2. Rime states that a minute is about 1,000 characters, so that is roughly $30 and $50 per 1M characters. A $100 per month Scale plan takes 20% off, at $0.024 and $0.040. Its Prism speech-to-speech model is $0.080 per minute. Self-hosting through Docker Compose or Kubernetes and a HIPAA BAA with SOC 2 Type II are listed for enterprise; custom voice cloning is Enterprise only.

Strengths. Among the cheapest mainstream voices, the strongest vendor latency claim, a self-host path, and timestamps for aligning captions or interruptions. Coda v2 is 42nd on the arena (1112). Weaknesses. The independent data reads 275 ms median and 360 ms at P95 for Mist v3, and 318 ms and 428 ms for Coda, far from the vendor's 37 ms. Mist v3 covers four languages, cloning requires an Enterprise contract, and we did not find a Mist v3 arena row in the AA data we read. Best for: English and Western European phone agents on a budget that want a self-hosting option later.

Hume Octave (shutting down)

Hume's Octave 2 launched on Oct 1, 2025 with 11 languages, a stated time to first audio under 200 ms, a claimed 40% speed gain and half the price of Octave 1, and voice cloning from a 15-second sample. Those are first-party launch claims, a year old. Hume has since changed direction. Cartesia's migration page and several third-party posts quote Hume's changelog as saying the TTS and EVI APIs end on Nov 13, 2026 at 12:01 a.m. EST, with account data deleted afterward. Hume's current homepage markets an audio decision model, and its developer docs index (llms.txt) lists no TTS product. We could not fetch the changelog page itself, so treat the exact date as third-party reported, and do not plan a new build on Octave.

Migration notes. Cartesia's guide says Hume's voice design and free-text acting directions do not carry over to Sonic. Artificial Analysis still lists Octave 2 at Elo 1058 (68th) and $87.50 per 1M characters, which shows it as a historical rather than a recommended option. If you depend on Octave voices, export what you need before the cut-off and audition replacements with your real scripts.

Open Models (Kokoro, Orpheus, Sesame CSM, Chatterbox)

Open weights let you run TTS on your own GPUs, keep audio inside your network, and pay for compute instead of characters. The price is engineering: you own streaming, batching, concurrency, monitoring and failover. If you go this route, our GPU cloud comparison covers where to rent the hardware. Figures below come from each project's repository or model card, fetched Nov 12, 2026, and project-reported numbers are labeled as such.

Kokoro 82M

Kokoro is the small, cheap option: 82 million parameters, Apache-2.0 weights, and the v1.0 release (Jan 27, 2025) covers 8 languages and 54 voices (English, Spanish, French, Hindi, Italian, Japanese, Brazilian Portuguese and Mandarin). It installs with pip install kokoro and needs espeak-ng. The model card puts the market rate under $1 per 1M characters, and Artificial Analysis lists $0.65 via Replicate, the cheapest entry in its 104-model table. It scores 1061 on the arena (66th), about level with Google's Chirp 3: HD. No voice cloning is documented. The repository has about 10,000 stars. Best for: cheap fallback voices, on-device or edge deployments, and prototypes.

Orpheus TTS

Orpheus is Canopy Labs' Apache-2.0 speech model built on a Llama backbone. The project reports about 200 ms streaming latency, or about 100 ms with input streaming (project-reported, not independently tested), and lists zero-shot voice cloning and emotion tags. A multilingual research preview appeared in April 2025, and Baseten is a listed hosting partner. The repository has about 6,400 stars and runs on vLLM (pip install orpheus-speech). Orpheus does not appear in the Artificial Analysis rows we read, so we have no arena score for it. Best for: teams that want cloning and emotion control without a vendor contract and can host a GPU service.

Sesame CSM-1B

Sesame's Conversational Speech Model is a 1B-parameter Llama backbone with an audio decoder that produces Mimi codes. The repository is Apache-2.0, needs a CUDA GPU, and has about 14,700 stars. Two limits matter for agents: it is audio generation only (it cannot write the text it speaks, so you still need an LLM), and Sesame says it is not expected to do well outside English. The checkpoint is gated on Hugging Face and the repository frames it as research and education. Best for: research and demos of conversational prosody, not a production default.

Chatterbox (retired)

Resemble AI's MIT-licensed Chatterbox became a popular open choice, with about 27,800 stars and variants including a 350M Turbo model for low-latency English voice agents with paralinguistic tags such as [laugh]. The README now carries a banner saying that, as of August 2026, Chatterbox is retired research, is no longer maintained, and Resemble AI does not sell TTS. Artificial Analysis still lists Chatterbox HD at 1110 (43rd) and the base model at 1021 (85th), but plan for no fixes or security updates. Best for: existing deployments. Do not start new work on it.

Open weights above the pack. Per Artificial Analysis, Fish Audio Drama 3 is the best open-weights model, at Elo 1274 (7th overall) and $15 per 1M characters hosted. We did not read its license, so check commercial terms before relying on it.

Also Worth a Look

The arena list has other models we did not cover in depth because we did not fetch full vendor documentation for them. Treat these as leads to test, not recommendations. Alibaba's Qwen-Audio-3.1-TTS-Plus ranks 3rd (1297, $19.31). Smallest.ai's Lightning V3.1 Pro ranks 8th (1272, $19.50). Inworld's Realtime TTS-2 ranks 9th (1266, $20.83) and its Flash variant is 17th (1213, $10.42), and the independent Cekura harness shows both Inworld models at the top of the time-to-first-audio table. Gradium leads that table at a 53.7 ms median. AssemblyAI-TTS-1 is 15th (1222, $15) and Voxtral 3 TTS is 18th (1213, $13) in the arena data. Soniox and xAI's Grok TTS appear in the Cekura latency data. If your shortlist comes up short on latency or price, these are the next names to audition.

Arena Benchmarks

The only public, neutral quality ranking we could read with live numbers is the Artificial Analysis text-to-speech leaderboard, fetched Nov 11, 2026 with 104 models. Elo is built from blind listener votes between pairs of clips, so it measures which voice people prefer, not how fast it starts or how well it reads a phone number. The table lists the models relevant to this guide. A rank gap of a few Elo points is a tie. For example, Eleven v4 at 1338 and v4 Turbo at 1337 are the same result in practice.

What the table says. Price and Elo are only loosely related. Simba 3.5 at $6.58 and Gemini 3.8 Flash-Lite at $11.03 both beat Cartesia and match or beat several models that cost four to eight times as much. At the other end, Polly Standard is last of 104 at 830 and Hume's Octave 2, at $87.50, sits at 1058. The legacy OpenAI models are mid-table. Cartesia's own August comparison cited Sonic 3.6 at 1282 against Eleven v3 at 1177, and the live table now shows 1296 and 1189, so the ordering held while the numbers shifted.

What we could not use. The Hugging Face TTS Arena page returned only an application shell when we fetched it on Nov 11, 2026, so we cite no numbers from it. Artificial Analysis also publishes a pronunciation-robustness benchmark, which we did not read in detail. Hume's own Real World VoiceEQ leaderboard (read through a third-party article on Nov 9, 2026, not from Hume directly) lists Gemini 3.8 Flash TTS at 0.92, Flash-Lite at 0.91, Eleven v4 Turbo at 0.91, v4 at 0.90 and Sonic 3.6 at 0.84. That is a vendor benchmark from a company that has since shut its TTS API, so we do not rank on it.

Latency: Claimed vs Measured

“Time to first audio” (TTFA) is the delay between sending text and receiving the first playable audio chunk. It decides how natural the pause feels after a caller stops talking, alongside speech recognition and the language model. Vendors measure it in their favor: a short text, one concurrent request, a nearby region, and often the model alone without the network. The independent figures below come from the Cekura harness as republished on voiceagent-tts.com (a third-party aggregator, checked Oct 27, 2026). We did not reproduce them.

Compare these to the vendor claims in the master table: Rime says 37 ms P50 for Mist v3 and the harness reads 275 ms; Cartesia says under 90 ms and the harness reads 286 ms; ElevenLabs says about 150 ms median time to first speech for v4 Turbo and the harness reads 189 ms. The ElevenLabs gap is the smallest, which suggests a conservative claim. The others differ by a factor of three to seven. The harness lists no latency for ElevenLabs v3 or v4, Gemini TTS, Hume Octave 2, or OpenAI's TTS models, so those rows are a gap, not a result.

How to read it. Tail latency matters more than the median in a phone call, because a rare 500 ms stall is what people notice. Deepgram Aura 2 is a clear case: 263 ms median but 525 ms at P95, a wide spread. A third-party guide (nextlevel.ai, not tested by us) also notes buffering defaults that change first-chunk timing: ElevenLabs' WebSocket API reportedly waits for 120 characters by default, and Cartesia's buffer maxes at 3,000 ms unless set to zero. Check those settings in the docs before you benchmark, because they can add or remove hundreds of milliseconds. If your voice stack already has a speech-to-text step, our speech-to-text comparison covers the other end of the latency budget.

A walkthrough of building a voice agent with Stream's Vision Agents and Cartesia Sonic 3, useful for seeing where TTS sits in a streaming pipeline.

Cost at Voice Agent Volume

List prices per character look small until you multiply by call volume. The table converts each option to a monthly bill at two volumes. We assume Rime's own conversion of about 1,000 characters per minute of speech, so 10M characters is about 10,000 minutes of agent speech (roughly 167 hours) and 100M is about 100,000 minutes. Real calls vary with speaking rate and language. Cartesia says a minute uses 750 to 800 credits, which would make its rows about 20% cheaper per minute than shown. Subscription minimums, free tiers and volume discounts are ignored, and Azure and Gemini use Artificial Analysis conversions rather than vendor list prices.

What it means. The spread between the cheapest mainstream voice and ElevenLabs v4 is about 2.7x ($300 versus $800 at 10M), and Turbo cuts that to 1.3x. If a human agent costs far more than the TTS bill, voice quality is a poor place to save money. If you are running a high-volume notification line, it is a very good place. Also count your LLM and STT costs: Deepgram's all-in Voice Agent API is $0.075 per minute, Cartesia's is $0.03 per minute, and ElevenLabs' Realtime v1 is $0.08 per minute plus LLM usage, all of which can exceed the TTS-only line items above. For another volume planning example, see how MoneyPrinterTurbo pairs generated narration with short video, a use case where batch TTS cost dominates.

How to Choose

Start from the constraint that cannot move: language, residency, latency budget or price per minute. Then audition two or three finalists with your own scripts.

If you want a managed stack that bundles TTS, STT and orchestration, compare voice agent platforms before you pick a standalone TTS API. A platform may lock the voice to its own bundled vendors or let you bring your own, and that can decide the choice for you. For a build walkthrough that wires a voice into a phone number, see building an AI phone calling agent. If the agent also answers from your own documents, see the vector database comparison for the retrieval layer that feeds the text you will be speaking. If you code with agents and want voice in your terminal, our Antigravity voice mode guide shows a related setup.

What to Test Before You Commit

1. Time to first audio from your region. Stream a 20 to 40 word reply from the server that will run the agent, for example 200 requests at your expected concurrency, and record median and P95. Do it twice at different hours.
2. Your hardest strings. Phone numbers, order IDs, dates, currency, email addresses, brand names, and any second language. Arena Elo does not test these. Artificial Analysis publishes a pronunciation-robustness benchmark if you want a starting point.
3. Streaming behavior. Feed text in as your LLM emits tokens and check whether the voice waits too long for a first chunk, changes tone between sentences, or clips the first syllable. Use the vendor's continuation or context feature if it has one.
4. Barge-in. Confirm you can cancel audio instantly when the caller interrupts, and that billing stops at the cancel.
5. Telephony audio. Request the narrowband format your carrier uses (for example ElevenLabs lists ulaw_8000) and listen on a real phone, not studio headphones.
6. Cloning terms. Read the consent, licensing and retention terms. Voice cloning without recorded consent is a legal and reputational risk, and OpenAI's usage policy requires disclosing AI-generated voices to end users.
7. Rate limits and concurrency. Self-serve plans cap concurrent streams; a third-party guide lists 2 to 15 for Cartesia from Free to Scale and 45 for Deepgram Aura-2 on pay-as-you-go. Verify against the vendor page and ask about burst limits.
8. A fallback voice. Wire a second provider behind a feature flag. Hume and OpenAI both gave customers months, not years, to migrate, and a cheap Polly or Aura-2 fallback also covers outages.

How We Evaluated

• Prices: each vendor's own pricing or docs page, fetched Nov 11 to 12, 2026. Per-minute, per-credit and per-token prices were converted to per 1M characters by our own arithmetic where noted. Azure's page did not render dollar values, so Azure uses Artificial Analysis figures and is labeled that way.
• Quality: the Artificial Analysis Speech Arena leaderboard, fetched Nov 11, 2026. The Hugging Face TTS Arena returned no readable data.
• Latency: vendor claims are labeled vendor-reported. The independent table is Cekura data via voiceagent-tts.com, checked Oct 27, 2026, and is a third-party measurement we did not reproduce.
• Open models: GitHub repositories and Hugging Face model cards, fetched Nov 12, 2026. Latency for Orpheus is project-reported.
• Hume and OpenAI shutdowns: OpenAI is from its own deprecations page and TTS guide. Hume's date is from Cartesia's migration page and other third parties quoting Hume's changelog; the changelog itself redirects to new product docs.
• Not tested by us: we did not run these APIs on our own audio. The ranking is a synthesis of published data, and no vendor paid for placement. There are no affiliate links in this article.
• Gaps we did not fill: vendor-published latency for Deepgram, Google, Azure, Polly and OpenAI; Azure per-1M list prices; a Mist v3 arena row; Hugging Face Arena scores; and a consistent language count for Azure.

FAQ

Sources

All pages fetched Nov 11 to 12, 2026 unless noted.

• Artificial Analysis: text-to-speech leaderboard (Elo, price, rank)
• voiceagent-tts.com: Cekura TTS latency benchmark (checked Oct 27, 2026)
• ElevenLabs: API pricing, Eleven v4 launch post and create speech API reference
• Cartesia: pricing, Sonic-3.6 launch post and TTS models docs
• Cartesia: migrating from Hume
• Deepgram: pricing and TTS models docs
• OpenAI: API pricing, deprecations and text-to-speech guide
• Google Cloud: Text-to-Speech pricing and Chirp 3: HD docs
• Microsoft Azure: Speech Services pricing
• Amazon Polly: pricing
• Rime: pricing
• Hume: pricing, developer docs index and Versely: Octave 2 article (third party)
• Kokoro on GitHub, Orpheus-TTS, Sesame CSM and Chatterbox
• Deepgram on YouTube: Best Text-to-Speech for Voice AI (embedded above) and The Best Text-to-Speech Platform for Voice AI (embedded above)