There is no universally best TTS provider. ElevenLabs leads on voice library depth and language breadth. Cartesia leads on language coverage with low latency. OpenAI is the natural choice for teams already in the OpenAI ecosystem. Gradium leads on TTS models for voice agents: robust pronunciation on the hard cases, voices optimized for live agent interactions, and robust voice cloning and custom voice design.
The right choice depends on what constraint actually determines whether your product works. This guide maps each provider's strengths to the use cases where those strengths matter.
What separates these TTS providers?
Before comparing providers, it helps to be clear about what each benchmark actually measures and what it does not.
Artificial Analysis ELO measures voice preference in blind pairwise human comparisons. It tells you which voice listeners preferred on a static English audio prompt. It does not measure latency, WER, multilingual performance, or how a voice holds up on structured content like phone numbers.
Coval TTS benchmark (benchmarks.coval.ai/tts) measures TTFA, latency IQR, and WER under production conditions across hundreds of runs per model. It captures how a TTS API behaves when it is actually serving traffic, not just responding to a single warm request. It only covers streaming-capable WebSocket APIs. Source: May 4, 2026 capture, 750 runs per model.
MiniMax Multilingual TTS Test Set measures WER per language across a public multilingual benchmark. Source: Gradium evaluation published April 29, 2026.
These three sources answer different questions. The right provider comparison depends on which question your use case is actually asking.
What does each provider do best, and where does each fall short?
Gradium
What it leads on: production latency (155 ms TTFA P50, lowest of 9 models on Coval), latency consistency (2 ms IQR, lowest on Coval), pronunciation accuracy (3.3% WER on Coval, lowest of 8 models with WER data), and multilingual WER (1.11% average across EN, FR, ES, PT, DE on MiniMax Multilingual benchmark).
Gradium's TTS is designed specifically for voice agents. As the comparison in What Is the Best TTS API in 2026 to Build Voice Agents concludes, no other provider matches Gradium across all five criteria that matter for production voice agents: latency, voice quality, pronunciation robustness, stability under load, and deployment flexibility. Raw text from an LLM can be sent directly to Gradium TTS without preprocessing for structured inputs. Phone numbers, dates, email addresses, URLs, and alphanumeric codes are handled natively. For domain-specific terms, pronunciation dictionaries are available.
STT with semantic VAD is included in the same platform and billing pool. Voice cloning (Instant, from 10 seconds of audio) is available from the free tier. In a blinded benchmark of 3,220 voice pairs across EN, FR, DE, and ES, Gradium's Instant Voice Clone achieved the highest Elo score in every language.
Deployment: cloud, private cloud, on-premise (HIPAA-compliant), and on-device (Phonon). Languages: EN, FR, DE, ES, PT.
Where it falls short: 5 supported languages vs 32 to 40+ for ElevenLabs and Cartesia. Voice quality ELO on Artificial Analysis is #24 (1,072), lower than the top-ranked providers. For teams whose product targets languages outside the five, or where the highest ELO voice quality is the only decision criterion, other providers rank above Gradium on those specific dimensions.
Best for: real-time voice agents, customer support automation, phone-based agents, and any application where TTFA, latency consistency, and pronunciation accuracy are the hard constraints.
ElevenLabs
ElevenLabs is the most widely recognized TTS platform for content creation and has the broadest commercial voice library.
What it leads on: voice library depth, language coverage (32 to 70+ depending on model), and Artificial Analysis ELO for content creation models (Eleven v3 at ELO 1,178, #4 globally).
The models relevant for real-time voice agents are Turbo v2.5 and Flash v2.5. On the Coval benchmark (May 4, 2026): Turbo v2.5 records 264 ms TTFA P50 (28 ms IQR, 5.2% WER), Flash v2.5 records 288 ms TTFA P50 (28 ms IQR, 5.2% WER). Both are above the 200 ms conversational threshold and record 5.2% WER vs Gradium's 3.3% on the same benchmark.
ElevenLabs Multilingual v2 records 1,232 ms TTFA P50 on Coval, making it unsuitable for real-time voice agents. It is appropriate for batch content generation and audiobooks.
STT is a separate product (Scribe v2), billed separately. Voice cloning is available on paid plans.
On the numbers in TTS Latency Benchmark 2026, Gradium TTS is 109 ms faster on P50 (155 ms vs 264 ms) and 14x more consistent on IQR (2 ms vs 28 ms), while ElevenLabs covers more languages (32 vs 5). For voice agents that require coverage beyond English, French, German, Spanish, and Portuguese, ElevenLabs Turbo v2.5 is currently the best option that still fits the real-time latency budget.
Pricing: Turbo v2.5 and Flash v2.5 at $50/1M characters. Eleven v3 at $100/1M.
Best for: content creation, audiobooks, dubbing, narration where TTFA is not the constraint. Voice agents requiring 32+ languages where Gradium's 5 do not cover the target market.
Not ideal for: applications where sub-200 ms TTFA is required, or where WER on structured content is the primary constraint.
Cartesia
Cartesia's Sonic line was built around low-latency streaming from the start, using State Space Model (SSM) architecture rather than standard transformers. The current releases are Sonic 3.5 (May 2026) and Sonic-3.6 (July 2026); the independent Coval data below covers Sonic-3, the version in the May 2026 capture.
What it leads on: TTFA among providers with broad language coverage. On the Coval benchmark, Sonic-3 records 188 ms TTFA P50, the second lowest of all 9 models, behind Gradium at 155 ms. Supports 40+ languages with documented regional accent variants, the widest coverage in this comparison. Cartesia Sonic-3 is the second option on Artificial Analysis (ELO 1,070, #25), one position behind Gradium (#24, ELO 1,072).
Where it falls short: 100 ms IQR on Coval is 50 times wider than Gradium's 2 ms. Per TTS Latency Benchmark 2026, Gradium TTS is 33 ms faster on P50 (155 ms vs 188 ms) and 50x more consistent on IQR (2 ms vs 100 ms). Cartesia Sonic-3 places second on raw P50 latency, but its 100 ms IQR means a significant fraction of requests cross the 300 ms conversational threshold. For voice agents where tail latency matters, Gradium's 2 ms IQR is the operationally safer choice.
WER data shows an anomaly in the Coval dataset for Cartesia, so no reliable WER figure is available for direct comparison.
Pricing: Startup at $39/1M (annual), Scale at $239/month (8M credits annual).
Best for: voice agents or content products requiring language coverage beyond Gradium's 5 languages, particularly with specific regional accent requirements (Latin American Spanish, Brazilian Portuguese, Bavarian German, and others).
OpenAI TTS
OpenAI offers TTS through its Audio API with three models: TTS-1 ($15/1M), TTS-1-HD ($30/1M), and GPT-4o Mini TTS (token-based pricing).
What it leads on: cost at $15/1M for TTS-1, the lowest per-character price among major hosted providers with documented quality rankings (ELO 1,102, #17 on Artificial Analysis, 7,548 samples, one of the most statistically robust rankings on the leaderboard). Convenience for teams already in the OpenAI ecosystem (single API key, billing, and SDK).
Where it falls short: OpenAI's standard TTS endpoint uses HTTP chunk transfer encoding, not WebSocket streaming. On the Coval benchmark, TTS-1-HD records 2,136 ms TTFA P50 with a 822 ms IQR, making it unsuitable for real-time voice agents. There is no publicly available self-serve voice cloning.
The OpenAI Realtime API handles speech-to-speech end-to-end over WebSocket, but it is not a modular TTS endpoint. It bundles STT, LLM reasoning, and TTS into a single model, which means individual components cannot be swapped.
Pricing: TTS-1 at $15/1M, TTS-1-HD at $30/1M.
Best for: batch content generation, narration in OpenAI-native stacks, and any use case where cost per character is the primary constraint and real-time streaming is not required.
Not ideal for: real-time voice agents where TTFA is a hard constraint.
Deepgram Aura-2
Deepgram is primarily an STT platform. Aura-2 is its TTS model, positioned for teams already using Deepgram Nova for transcription.
What it leads on: a unified STT + TTS platform where Deepgram Nova-3 handles transcription and Aura-2 handles synthesis, reducing vendor fragmentation for teams already on Deepgram. 7 languages.
Where it falls short: Aura-2 records 313 ms TTFA P50 on Coval, the slowest viable real-time model in the benchmark, with 6.4% WER, the highest of any real-time-viable provider. No voice cloning available.
Pricing: $30/1M characters.
Best for: teams already using Deepgram Nova for STT who want TTS in the same platform without adding a new vendor.
Google Gemini 3.1 Flash TTS and Inworld Realtime TTS 1.5 Max
Two providers that lead on Artificial Analysis ELO but are not in the Coval production benchmark.
Google Gemini 3.1 Flash TTS: ELO 1,206 (#2 on Artificial Analysis, 1,890 samples), $36.6/1M. Strong choice for teams on Google Cloud infrastructure where voice quality ELO is the primary criterion.
Inworld Realtime TTS 1.5 Max: ELO 1,208 (#1 on Artificial Analysis, 1,851 samples), $35/1M. The highest human-rated TTS model currently on the market at a competitive price. Includes zero-shot voice cloning and a Realtime API. 15 languages. No independent TTFA data available from Coval.
Neither provider is in the Coval production benchmark, so independent TTFA and WER data under the same production conditions as Gradium, Cartesia, ElevenLabs, and Deepgram are not available for direct comparison.
Which provider fits which constraint?
| If your primary constraint is | Choose |
|---|---|
| TTFA under 200 ms + low WER | Gradium |
| Language coverage beyond 5 languages, low latency | Cartesia Sonic-3 |
| Highest voice quality ELO | Inworld TTS 1.5 Max |
| Second-highest ELO + Google Cloud | Google Gemini 3.1 Flash TTS |
| Broad language coverage (32-70+) for content | ElevenLabs |
| Lowest per-character cost | OpenAI TTS-1 |
| Single vendor STT + TTS (already on Deepgram) | Deepgram Aura-2 |
| Voice cloning from free tier | Gradium |
| On-premise / on-device deployment | Gradium |
| Open-source / self-hostable | Fish Audio S2 Pro (Apache 2.0) |
What should you ask before picking a TTS provider?
What is your TTFA requirement? If sub-200 ms is a hard constraint, only Gradium (155 ms) and Cartesia (188 ms) on the Coval benchmark consistently meet it at the median. If TTFA is flexible (content creation, batch generation), the full provider set is available.
How many languages do you need? Gradium covers 5 (EN, FR, DE, ES, PT) with native fluency and mid-sentence code-switching. Cartesia covers 40+, ElevenLabs 32 to 70+. If your product targets markets outside Gradium's five, Cartesia or ElevenLabs is the starting point.
Do you need voice cloning? Gradium offers Instant Voice Cloning from the free tier (10s of audio, no credit card). ElevenLabs offers it on paid plans. Cartesia on paid plans. Deepgram does not offer voice cloning.
Is STT included or separate? Gradium's STT with semantic VAD shares the same credit pool as TTS. ElevenLabs STT (Scribe v2) is a separate product. Deepgram covers both but as separate billing lines.
What is your deployment constraint? If cloud-only is fine, all providers apply. If on-premise or on-device is required, Gradium is the only provider in this comparison with all four deployment options (cloud, private cloud, on-premise, on-device via Phonon).
Glossary
Artificial Analysis ELO
A score derived from blind pairwise human comparisons of TTS audio samples. Higher ELO means more evaluators preferred that voice in blind tests. Does not measure TTFA, WER, multilingual performance, or streaming behavior. Source: artificialanalysis.ai.
Coval TTS benchmark
An independent production benchmark (benchmarks.coval.ai/tts) measuring TTFA P50, latency IQR, and WER across streaming WebSocket TTS APIs under production conditions. Open-source methodology. Data referenced from May 4, 2026 capture.
TTFA (Time to First Audio)
The elapsed time between sending text to a TTS API and receiving the first streamed audio chunk. Primary latency metric for real-time voice agents. Gradium: 155 ms P50 (Coval, May 4, 2026). ElevenLabs Turbo v2.5: 264 ms. Cartesia Sonic-3: 188 ms. OpenAI TTS-1-HD: 2,136 ms.
Latency IQR
The Interquartile Range between P25 and P75 TTFA values. Measures latency consistency across production requests. Gradium: 2 ms (near-deterministic). Cartesia Sonic-3: 100 ms. ElevenLabs Turbo v2.5: 28 ms.
Word Error Rate (WER) for TTS
Pronunciation accuracy measured by transcribing synthesized audio and comparing to input text. Gradium: 3.3% on Coval (lowest of 8 models with data), 1.11% on MiniMax Multilingual. ElevenLabs Turbo/Flash: 5.2% Coval. Deepgram Aura-2: 6.4% Coval.