1
ElevenLabs Eleven v4
ElevenLabs · All-round narration, character voices, cloning and dubbing in one app
$$$
Eleven v4 Turbo and Eleven v4 hold the top two places on the Artificial Analysis Provider Voice Arena (Elo 1336 and 1334). Released 28 September 2026, v4 brings back Professional Voice Clones, follows inline audio tags and direction prompts, and handles multi-speaker scenes. It sits inside the most complete creator toolkit: voice changer, dubbing, sound effects and music under one subscription.
- Cheapest way in
- Free plan (personal use); Starter $6/month with commercial licence
- Cost per output
- $0.08 per 1K characters via API ($0.04 on v4 Turbo); dubbing $0.33-0.50 per minute
- Languages
- 90+
- Voice cloning
- Instant (about 10s of audio) and Professional
- Emotion control
- Inline audio tags and natural-language direction
- Latency
- v4 Turbo about 150ms median time to first speech (vendor figure)
- Open weights
- No
- Multiple speakers
- Yes
Good at
- Top of the main blind listening arena
- 90+ languages with a cloned voice carried across languages
- Instant clones from about 10 seconds of audio plus Professional Voice Clones
- Inline audio tags such as [whispering] or [laughs] for performance control
Watch out for
- Dearer per character than Google, Fish Audio or Inworld at API level
- Free plan is personal use only
- Benchmark claims in the launch post are vendor-reported without links
Benchmarks Artificial Analysis Provider Voice Arena (Text to Speech): #1 Eleven v4 Turbo, Elo 1336; #2 Eleven v4, Elo 1334
Prices elevenlabs.io, checked 1 November 2026
ElevenLabs Eleven v4 website Work made with it on AI LIVE
2
Gemini 3.8 Flash TTS
Google · Cheap, high-quality multi-speaker dialogue and voiceover at scale
$
Gemini 3.8 Flash TTS is #5 on the Artificial Analysis Provider Voice Arena (Elo 1283), with Flash-Lite at #10 (Elo 1259), at a fraction of ElevenLabs' price. Released 23 September 2026, it follows per-line stage directions, offers 2,000+ preset voices and voice design from a description, and has a free tier in Google AI Studio. Voice cloning exists but needs recorded consent and is blocked in the UK and EEA, which matters for UK creators.
- Cheapest way in
- Free tier in the Gemini API / Google AI Studio
- Cost per output
- About $0.0135 per minute of audio (Flash TTS) until 31 Dec 2026; prices double from 1 Jan 2027
- Languages
- 100+
- Voice cloning
- From a 30s sample with consent; blocked in the UK and EEA
- Emotion control
- Per-line stage directions
- Open weights
- No
- Voices
- 2,000+ preset plus voice design from text
- Multiple speakers
- Yes
Good at
- Near the top of the arena at very low cost
- Free tier in the Gemini API and Google AI Studio
- Line-by-line delivery direction and multi-speaker dialogue
- 100+ languages and 2,000+ voices
Watch out for
- Voice cloning blocked in the UK, EEA, Switzerland, India, Illinois and Texas
- Introductory pricing doubles on 1 January 2027
- Developer-oriented: no dedicated creator editing app
Benchmarks Artificial Analysis Provider Voice Arena (Text to Speech): #5 Gemini 3.8 Flash TTS, Elo 1283; #10 Gemini 3.8 Flash-Lite TTS, Elo 1259
Prices ai.google.dev, checked 1 November 2026
Gemini 3.8 Flash TTS website
3
Fish Audio Drama 3
Fish Audio · Directed multi-character drama, and the best open-weights voice model
$$
Drama 3 is #6 on the Artificial Analysis Provider Voice Arena (Elo 1278) and the highest-ranked open-weights model there. Final release was 13 October 2026: a 35B mixture-of-experts model that voices up to 16 characters in one scene, takes bracketed acting direction and covers 83 languages. Weights are on Hugging Face under a community licence that is free for commercial use below US$500,000 annual revenue.
- Cheapest way in
- Free plan (personal use); Plus from $11/month with commercial use
- Cost per output
- $10 per 1M UTF-8 bytes via API until 12 Nov 2026, then $50
- Languages
- 83 (vendor figure)
- Voice cloning
- Yes, all plans; cloning a real person without consent banned by licence
- Emotion control
- Bracketed direction, paired tags such as <whisper>, placed breaths
- Open weights
- Yes, community licence (free commercial use under US$500k revenue)
- Multiple speakers
- Up to 16 characters
- Parameters
- 35B MoE, about 3B active
Good at
- Open weights (fishaudio/drama-3) for self-hosting
- Up to 16 characters per scene with inline direction tags
- Strong pronunciation (93.2% on Artificial Analysis pronunciation test, vendor-reported)
- Cheap web plans with cloning on every tier
Watch out for
- Launch API price of $10 per M bytes rises to $50 after 12 November 2026
- Licence requires a paid deal above US$500,000 revenue or to resell as a voice API
- Self-hosting needs roughly 32-40 GB of VRAM
Benchmarks Artificial Analysis Provider Voice Arena (Text to Speech): #6, Elo 1278 (top open-weights model on the board)
Prices docs.fish.audio, checked 1 November 2026
Fish Audio Drama 3 website Work made with it on AI LIVE
4
Cartesia Sonic 3.6
Cartesia · Fast, natural voiceover and accent-faithful instant clones
$$
Sonic 3.6 is #4 on the Artificial Analysis Provider Voice Arena (Elo 1296), released 27 August 2026. It covers 44 languages with 500+ preset voices and improved accent retention in instant clones, and replies in under 90ms. It is built mainly for real-time voice agents, so creative controls are lighter than ElevenLabs or Fish Audio.
- Cheapest way in
- Free plan; Pro $5/month with commercial use and instant cloning
- Cost per output
- About $50 per 1M characters on Pro, $37-39 on higher plans; overage $65 per 1M
- Languages
- 44
- Voice cloning
- Instant (Pro and up), Professional (Startup and up)
- Emotion control
- Context-adaptive delivery; no explicit controls described
- Latency
- Under 90ms (vendor figure)
- Open weights
- No
- Voices
- 500+ preset
Good at
- High arena score
- Very low latency and fast generation
- Cheap entry plan ($5/month) with commercial use and instant cloning
- Hinglish and Indic locale support
Watch out for
- Launch post describes no explicit emotion or direction controls
- Professional cloning only from the $49 Startup plan
- No dubbing or editing studio for creators
Benchmarks Artificial Analysis Provider Voice Arena (Text to Speech): #4, Elo 1296
Prices cartesia.ai, checked 1 November 2026
Cartesia Sonic 3.6 website
5
Qwen-Audio 3.1 TTS
Alibaba · High-quality TTS for developers already on Alibaba Cloud
$
Qwen-Audio 3.1 TTS Plus is #3 on the Artificial Analysis Provider Voice Arena (Elo 1305), the highest non-ElevenLabs score. It was released around 23 September 2026 as part of a five-model speech family, hosted only in Alibaba Cloud Model Studio. It ranks below its Elo because access is API-only with no creator app and no open weights.
- Cheapest way in
- Pay as you go via Alibaba Cloud Model Studio
- Cost per output
- About $19.30 per 1M characters (Artificial Analysis listing)
- Open weights
- No
- Access
- Alibaba Cloud Model Studio API (model ID qwen-audio-3.1-tts-plus)
- Tiers
- Plus and Flash
Good at
- Third-best voice quality on the arena
- Mid-range price per character
- Alibaba reports strong cross-language voice similarity (vendor charts)
Watch out for
- Hosted only via Alibaba Cloud Model Studio; no open weights
- No consumer editing app
- Alibaba's official pricing and language list are unconfirmed
Benchmarks Artificial Analysis Provider Voice Arena (Text to Speech): #3, Elo 1305
Prices artificialanalysis.ai, checked 1 November 2026
Qwen-Audio 3.1 TTS website
6
Inworld TTS-2
Inworld · Budget voiceover with cloning and voice design on a free plan
$
Realtime TTS-2 is #8 on the Artificial Analysis Provider Voice Arena (Elo 1267), with the cheaper TTS-2 Flash at #16 (Elo 1217). The free on-demand plan includes about 70 minutes of TTS, instant cloning from 5-15 seconds of audio, professional clones and voice design from text. Pricing is among the lowest of the top ten.
- Cheapest way in
- Free plan with up to 70 minutes of TTS; Creator $25/month
- Cost per output
- Realtime TTS-2 $25 per 1M characters on demand ($20 on Creator); TTS-2 Flash $15 per 1M ($10 on Creator)
- Languages
- 200+ (Inworld's claim)
- Voice cloning
- Instant (5-15s) and Professional
- Voice design
- Yes
- Latency
- Under 100ms p99 (vendor figure)
- Open weights
- No
Good at
- Generous free tier with cloning and voice design
- Low per-character cost
- Steering controls on Realtime TTS-2
Watch out for
- Aimed at games and voice agents more than long-form narration
- No dubbing workflow
- The 200+ language count is Inworld's own claim
Benchmarks Artificial Analysis Provider Voice Arena (Text to Speech): #8 Realtime TTS-2, Elo 1267; #16 TTS-2 Flash, Elo 1217
Prices inworld.ai, checked 1 November 2026
Inworld TTS-2 website
7
MiniMax Speech 2.8
MiniMax · Quick voice clones for characters across many languages
$$$
Speech 2.8 HD is #27 on the Artificial Analysis Provider Voice Arena (Elo 1174), behind the leaders but widely available through MiniMax Audio and third-party hosts such as Replicate and Cloudflare. Rapid cloning from about 10 seconds of audio costs $1.50 per voice. Per-character pricing is the highest in this list.
- Cheapest way in
- Pay as you go via API; also in the MiniMax Audio app
- Cost per output
- $100 per 1M characters (HD), $60 (Turbo); voice cloning $1.50 per voice
- Languages
- 32 to 40+ (sources disagree)
- Voice cloning
- Rapid cloning from about 10s
- Voice design
- Yes
- Open weights
- No
Good at
- Fast cloning from short clips
- Turbo variant for cheaper drafts
- Available on many creative platforms
Watch out for
- Expensive per character at $100 per 1M for HD
- Well behind 2026 leaders on blind listening
- Language count varies by source (32 to 40+)
Benchmarks Artificial Analysis Provider Voice Arena (Text to Speech): #27 Speech 2.8 HD, Elo 1174; #31 Speech 2.8 Turbo, Elo 1153
Prices platform.minimax.io, checked 1 November 2026
MiniMax Speech 2.8 website
8
Mistral Voxtral 3 TTS
Mistral AI · Low-cost European TTS, with an older open-weights option
$
Voxtral 3 TTS is #15 on the Artificial Analysis Provider Voice Arena (Elo 1223) at about $13 per 1M characters. Voxtral 3 (October 2026) appears to be API-only; the earlier March 2026 Voxtral TTS (4B) has open weights, reportedly under a non-commercial CC BY-NC 4.0 licence.
- Cheapest way in
- Pay-as-you-go via Mistral API
- Cost per output
- About $13 per 1M characters (Artificial Analysis listing)
- Open weights
- Voxtral 3: no. Voxtral TTS (Mar 2026, 4B): yes, reportedly CC BY-NC 4.0
Good at
- Good arena score for the price
- Older 4B model available as open weights for experiments
Watch out for
- Voxtral 3 weights not published
- Open-weights licence reportedly non-commercial
- Little official detail on the new model
Benchmarks Artificial Analysis Provider Voice Arena (Text to Speech): #15, Elo 1223
Prices artificialanalysis.ai, checked 1 November 2026
Mistral Voxtral 3 TTS website
9
Hume Octave 2
Hume AI · Designing a voice from a written description and acting it out
$$$
Octave 2 (still labelled preview) is built around voice design from natural-language prompts plus acting instructions for delivery, which suits character work. Its blind-test score is mid-table: #63 on the Artificial Analysis Provider Voice Arena (Elo 1060). It supports 11 languages and voice cloning; Free and Starter plans are non-commercial.
- Cheapest way in
- Free plan (non-commercial use only)
- Cost per output
- About $87.50 per 1M characters (Artificial Analysis listing)
- Languages
- 11, including English, Japanese, Spanish, Hindi and Arabic
- Voice cloning
- Yes
- Emotion control
- Acting instructions, speed, pitch, emotion
- Voice design
- Yes
- Open weights
- No
Good at
- Voice design from a text description
- Acting instructions, speed, pitch and emotion controls
- Voice cloning supported
Watch out for
- Low arena placing for 2026
- Still a preview
- Free and Starter plans are non-commercial
Benchmarks Artificial Analysis Provider Voice Arena (Text to Speech): #63, Elo 1060
Prices dev.hume.ai, checked 1 November 2026
Hume Octave 2 website
10
Resemble Chatterbox
Resemble AI · Free, MIT-licensed voice cloning you can run yourself
$
Chatterbox is the most permissively licensed option here: the open model family is MIT-licensed, so commercial self-hosting is allowed without revenue caps. It offers zero-shot cloning with an emotion-intensity control, and a multilingual version covering 23 languages. Resemble's hosted Chatterbox HD sits at #39 on the Artificial Analysis Provider Voice Arena (Elo 1110), so quality trails the leaders.
- Cheapest way in
- Free (open weights, MIT)
- Cost per output
- Hosted Chatterbox HD about $40 per 1M characters (Artificial Analysis listing)
- Languages
- 23 (Chatterbox Multilingual)
- Voice cloning
- Zero-shot
- Emotion control
- Emotion exaggeration slider
- Open weights
- Yes, MIT
Good at
- MIT licence: free commercial self-hosting
- Zero-shot cloning from short clips
- Emotion exaggeration control
Watch out for
- Well behind 2026 closed models in blind tests
- Self-hosting needs a GPU and some technical setup
- Hosted TTS pricing not clearly published by Resemble
Benchmarks Artificial Analysis Provider Voice Arena (Text to Speech): #39 Chatterbox HD, Elo 1110; #80 Chatterbox, Elo 1024
Prices artificialanalysis.ai, checked 1 November 2026
Resemble Chatterbox website
