The text-to-speech leaderboard was rewritten this autumn. Between August and September, Eleven v4, Qwen-Audio 3.1 TTS-Plus, Cartesia Sonic 3.6 and Gemini 3.8 Flash TTS took the top five places on Artificial Analysis’s Speech Arena, and ElevenLabs’ own previous flagship, Eleven v3, fell to #18. The most useful number in this ranking is price: the cheapest model in the top five (Gemini 3.8 Flash TTS, $16.50 per 1M characters) costs about a fifth of the dearest (Eleven v4, $80) for a score only 52 Elo points lower. And Eleven v4 Turbo, at half v4’s price, scores the same as v4. So the question is less who sounds best than what the last few points are worth to you.
Five picks, by job — with per-character math throughout.
How we picked
Same rules as every ranking on this site: sourced claims, dated pricing, models you can actually buy. Rank comes from Artificial Analysis’s blind listener votes on two boards. On the provider voices board each model reads its own eight native voices; on the controlled voice board every model reads the same eight cloned voices, which isolates cloning quality. Scores carry roughly ±15 Elo of error, so we treat gaps inside that as ties and break them on price and access.
The two boards use separate scales, so compare scores within a board, not across them. Prices are Artificial Analysis’s cost to generate 1M characters at each model’s default settings, cross-checked against vendor price lists.
Two picks sit outside the arena for different reasons. Amazon Nova 2.5 Sonic is a speech-to-speech model that listens as well as speaks, so it has no place on a text-to-speech board; we place it by what it does. MAI-Voice-2.1 is in public preview and has not been measured, so it ranks last until independent scores exist.
The short answer
| Your situation | Pick |
|---|---|
| Best quality, and the default for most projects | Eleven v4 |
| Near-top quality at a quarter of the price | Qwen-Audio 3.1 |
| Designing voices from a prompt, long-form narration, 130 languages | Gemini 3.8 Flash TTS |
| A full voice-agent loop on AWS | Nova 2.5 Sonic |
| An all-Microsoft stack, willing to run a preview | MAI-Voice-2.1 |
All five, side by side
A fivefold price spread among the top models, in one picture — list price per 1M characters (MiniMax Speech 2.8 HD added for reference):
The full decision matrix — scores and prices from Artificial Analysis, specs from each model’s page:
| Model | Voice Elo (rank) | Cloned-voice Elo (rank) | Price / 1M chars | Languages | Access |
|---|---|---|---|---|---|
| Eleven v4 (Turbo / standard) | 1330 (#1) / 1326 (#2) | 1166 (#3) / 1161 (#4) | $40 / $80 | 90+ | ElevenAPI, free plan |
| Qwen-Audio 3.1 (TTS-Plus) | 1298 (#3) | 1185 (#2) | $19.30 | — | Alibaba Cloud API only |
| Gemini 3.8 Flash TTS | 1274 (#5) | 1050 (#17) | $16.50 | 130 | Gemini API, free tier |
| Nova 2.5 Sonic | not on board | not on board | ~$3 / $12 per 1M speech tokens | 7 | Bedrock, four regions |
| MAI-Voice-2.1 | not on board | not on board | $22 ($15 Flash) | 23 | Azure, public preview |
Dashes mean the vendor doesn’t state it plainly.
The reasoning, pick by pick
1. Eleven v4 — the quality leader
Eleven v4 is two models, and both lead. Eleven v4 Turbo is #1 on the provider-voices board (Elo 1330) and Eleven v4 is #2 (1326); the four-point gap is far inside the error bars. On the cloned-voice board they are #3 and #4 (1166 and 1161), behind Qwen-Audio 3.1 TTS-Plus and HeyGen Voice. Both take inline audio tags such as [laughs] and [excited, happy] and natural-language direction, run multi-speaker dialogue, support 90+ languages and clone an Instant Voice from about 10 seconds of audio.
The practical advice is to start with Turbo. ElevenLabs lists it at $0.04 per 1,000 characters against $0.08 for v4, reports about 100 ms of median inference, and the blind votes cannot tell them apart. Choose v4 for long produced work: it takes 10,000 characters per request and is tuned for audiobooks, voice-over and dubbing. ElevenLabs also reports that about 75% of listeners preferred v4 in its own blind tests against four rivals; that is a vendor-run test we have not verified.
Strengths: #1 and #2 on blind votes; audio tags and dialogue; 90+ languages; strong cloning; Turbo for agents. Watch out: the dearest leader at $80 per 1M characters for v4; closed and hosted only; cloning trails Qwen-Audio 3.1 on the controlled-voice board.
2. Qwen-Audio 3.1 — the value pick
Qwen-Audio 3.1’s TTS-Plus tier is #3 on provider voices (Elo 1298), 24 points ahead of Gemini 3.8 Flash TTS, and #2 on the cloned-voice board (1185), ahead of both Eleven v4 models. It costs $19.30 per 1M characters on Artificial Analysis’s figures — a quarter of Eleven v4 and about half of Turbo. The family around it is broad: speech recognition, a Chinese-and-English unified audio generator (TTS-Next) and a full-duplex realtime model, all through Alibaba Cloud.
One caveat is important. The ranked entry is a character-billed TTS-Plus tier, not the token-billed TTS-Flash on our model page, and we could not find an Alibaba Model Studio page for a 3.1 Plus model ID. The documented Plus model is qwen-audio-3.0-tts-plus, listed at $0.20 per 10,000 characters internationally and $0.19253 in Beijing, which works out to the $19.30 per 1M Artificial Analysis lists. It ranks #6 (1267), within the error bars of Gemini. Check which model ID you get before building on the #3 number.
Strengths: #3 on blind votes and #2 for cloned voices; $19.30 per 1M characters; ASR and realtime in the same family; deep Chinese dialect coverage. Watch out: API-only; ranked 3.1 Plus ID not in public docs; Alibaba’s published benchmarks are its own.
3. Gemini 3.8 Flash TTS — voice design and long-form
Gemini 3.8 Flash TTS is #5 on provider voices (1274) at $16.50 per 1M characters, and its Flash-Lite sibling is #8 (1246) at $11.00. The case for it is control rather than the top score: it designs a new voice from a text prompt, offers 2,000+ ready-made voices, covers 130 languages, directs each line through structured metadata, stages two-speaker scenes and keeps a voice steady across long recordings. There is a free tier on the Gemini API.
Two cautions. On the cloned-voice board it falls to #17 (1050), so cloning a specific voice is not its strength. And $16.50 is consistent with Google’s introductory rates, which double on January 1, 2027: budget for about twice that from next year. It is also a generator, not an agent — no Live API, no function calling.
Strengths: prompt-based voice design; 130 languages; long-form stability; cheapest of the top five; free tier. Watch out: #17 for cloned voices; price doubles in 2027; no live conversation; prompts written for older Gemini TTS need migration.
4. Amazon Nova 2.5 Sonic — voice agents on AWS
Nova 2.5 Sonic is on this list because many people searching for a voice generator are building a call agent, and for that job a text-to-speech engine is only a third of the loop. Nova Sonic is a single speech-to-speech model: it hears the caller, reasons, calls tools and speaks in one stream, with controllable turn-taking and asynchronous tool calls, on Amazon Bedrock. It is the same price as Nova 2 Sonic — tracked at about $3 per 1M speech input tokens and $12 per 1M speech output tokens, to be confirmed in the Bedrock console.
The limits are real. AWS publishes no benchmark or latency numbers for 2.5, supports seven languages, offers it in four regions and caps each connection at eight minutes, so long calls need restarts. It cannot be compared with the TTS boards, so we place it by fit: the pick for teams that already live on Bedrock.
Strengths: one model for hearing, reasoning and speaking; tool calling; Bedrock security and billing; Strands Bidi Agents. Watch out: no 2.5 benchmarks; seven languages; four regions; eight-minute connections.
5. MAI-Voice-2.1 — Azure’s preview
MAI-Voice-2.1 is Microsoft AI’s text-to-speech pair, announced on October 1: the main model at $22 per 1M characters and a Flash tier at $15 with about 45 ms of inference. Its distinctive claim is one voice identity across 23 languages with a native accent in each, plus gated cloning from a 5–60 second clip and SSML emotion control, delivered through Azure Speech and Foundry.
Everything about it is provisional. It is in public preview with no SLA and Microsoft’s docs do not recommend it for production. Neither model is on the Speech Arena yet, and the earlier MAI-Voice-1 sits at #72 (Elo 1033), so no one outside Microsoft has measured how good the new generation is. It earns a slot as the option for Azure teams to test, and it should move once real scores appear.
Strengths: one voice across 23 languages; Flash tier at about 45 ms and $15; Azure integration; consent guardrails on cloning. Watch out: public preview, no SLA; unmeasured; cloning is gated.
Also considered
These are close contenders or were replaced in this ranking. None made the five:
- Cartesia Sonic 3.6 — #4 on provider voices (1286) and #5 for cloned voices (1144) at $49.00 per 1M characters. Within the error bars of Qwen-Audio 3.1 and Gemini on quality, but at two to three times their price. We have no model page for it yet.
- Realtime TTS-2 — #7 (1260) at $20.80, and #6 for cloned voices (1143). Priced a little above Gemini and scored slightly lower; we do not cover it on a model page yet.
- HeyGen Voice — #1 on the cloned-voice board (1203) at $30.00, released in October 2026, but not on the provider-voices board. If cloning is your only need, test it.
- Gemini 3.8 Flash-Lite TTS — #8 (1246) at $11.00; the cheaper sibling of the Flash pick, covered on the Gemini 3.8 Flash TTS page.
- Simba 3.2 — #9 (1244) at $6.60, the cheapest model in the top ten by a wide margin.
- Breeze TTS 2 — #10 (1223) at $34.00 and the highest-ranked open-weights model on Artificial Analysis’s board, the starting point for self-hosting.
- ElevenLabs v3 — #18 (1177); replaced by Eleven v4 at the same $0.08 per 1,000 characters list price.
- MiniMax Speech 2.8 — #19 (1174, HD) and #24 (1153, Turbo) at $100 and $60 per 1M characters. Its remaining edge is $1.50 per cloned voice and $3 per designed voice.
- Kokoro 82M — the cheapest model on the board at $0.65 per 1M characters, open weights, #55 (1063). The budget route when quality matters less.
Looking for music instead? See the audio models directory. Full specs for every voice model live there too. The ranking was last reviewed on October 13, 2026.
Changelog
- 2026-10-13 — First version, built on Artificial Analysis’s provider-voices and controlled-voice Speech Arena boards. Ranking: Eleven v4, Qwen-Audio 3.1, Gemini 3.8 Flash TTS, Nova 2.5 Sonic, MAI-Voice-2.1. ElevenLabs v3 and MiniMax Speech 2.8 sit in “Also considered”.
Frequently asked questions
What is the best AI voice generator right now?
Eleven v4 from ElevenLabs. Eleven v4 Turbo is #1 on the Artificial Analysis Speech Arena (Elo 1330) and Eleven v4 is #2 (1326), with a gap that sits inside the error bars. If you want the best quality per dollar, Qwen-Audio 3.1 TTS-Plus is #3 at $19.30 per 1M characters, about a quarter of Eleven v4's $80.
Is Eleven v4 worth the price?
Mostly its Turbo variant is. Eleven v4 Turbo scores the same as Eleven v4 in blind votes (1330 vs 1326) for half the list price ($0.04 versus $0.08 per 1,000 characters). Pay for v4 only if you need its produced-content tuning and 10,000-character requests. If cost matters more than the last 30–60 Elo points, Qwen-Audio 3.1 and Gemini 3.8 Flash TTS cost $16.50–$19.30 per 1M characters.
What is the cheapest good AI voice generator?
Among the top five, Gemini 3.8 Flash TTS at $16.50 per 1M characters (#5). A little lower down are Gemini 3.8 Flash-Lite TTS at $11.00 (#8, Elo 1246) and Simba 3.2 at $6.60 (#9, 1244), and the open-weights Kokoro 82M at $0.65 (#55, 1063) is the cheapest model on the board. Google's prices double on January 1, 2027.
Which AI voice generator is best for voice cloning?
On Artificial Analysis's cloned-voice board, where every model reads the same eight cloned voices, HeyGen Voice is #1 (Elo 1203, $30 per 1M characters), Qwen-Audio 3.1 TTS-Plus #2 (1185), Eleven v4 Turbo #3 (1166) and Eleven v4 #4 (1161). Gemini 3.8 Flash TTS drops to #17 (1050). MiniMax Speech charges $1.50 per cloned voice, but ranks #20–21 there.
Which AI voice generator is best for voice agents?
For a text-to-speech layer, Eleven v4 Turbo: it ranks first and Eleven reports about 100 ms of median inference. For an all-in-one speech-to-speech agent on AWS, Amazon Nova 2.5 Sonic; Qwen-Audio 3.1 Realtime-Plus is a full-duplex alternative. MAI-Voice-2.1-Flash targets agents at about 45 ms inference but is still in public preview.
Are there open-weights AI voice generators?
Yes, but none of our five picks are open. Artificial Analysis lists Breeze TTS 2 as the top open-weights model at #10 (Elo 1223), and Kokoro 82M as the cheapest at $0.65 per 1M characters. We do not cover either on a model page yet.
What happened to ElevenLabs v3 and MiniMax Speech?
Both were overtaken. Eleven v3 is now #18 on blind votes (1177) and MiniMax Speech 2.8 HD is #19 (1174), at $100 per 1M characters on Artificial Analysis's figures. Eleven v4 replaced v3 on September 28, 2026, and MiniMax's $1.50 per-voice cloning fee is its remaining edge.
How are these models ranked?
By Artificial Analysis's blind listener votes on two Speech Arena boards: provider voices, where each model reads its own eight native voices, and controlled voice, where every model reads the same eight cloned voices. We treat gaps inside the roughly ±15 Elo error bars as ties and break them on price and access. Nova 2.5 Sonic (speech-to-speech) and MAI-Voice-2.1 (unmeasured) are placed by documentation, not votes.
How often is this ranking updated?
We re-check picks when major models ship. This first version was reviewed on October 13, 2026 against Artificial Analysis's two Speech Arena boards, vendor price lists and official documentation.
Sources
- Artificial Analysis — Text to Speech leaderboard (provider voices)
- Artificial Analysis — Text to Speech leaderboard (controlled voice)
- ElevenLabs — Introducing Eleven v4
- ElevenLabs — ElevenAPI pricing
- Alibaba Cloud Model Studio — qwen-audio-3.0-tts-plus
- Alibaba Cloud Model Studio — qwen-audio-3.1-tts-flash
- Google — Gemini 3.8 Flash TTS and Flash-Lite TTS
- Gemini API pricing
- AWS — Announcing Amazon Nova 2.5 Sonic
- Microsoft AI — MAI-Voice-2.1
- Microsoft Learn — MAI-Voice-2.1 and MAI-Voice-2.1-Flash
- MiniMax API Docs — pay-as-you-go pricing