Audio Models · checked Oct 17, 2026
Hear the leaderboard.
Voice, music and sound-effect models ranked by Artificial Analysis arenas, with list prices, per-minute estimates and what runs locally. Every number links to its source.
Leaderboards
Each board keeps its own scale. Speech shows the top 25 of 97 models on the board.
Text-to-Speech: Artificial Analysis
Swipe the chart sideways for all 25 models →
Qwen-
Gemini 3.8 Flash TTSQwen-Gemini 3.8 Flash-Gemini 3.1 Flash TTSElo (Artificial Analysis arena) · 25 models · top 5 coloured · axis does not start at zeroSource: artificialanalysis.ai ↗ · checked 2026-10-23
Music with vocals: AA-Music-Vocal v1.1
Swipe the chart sideways for all 13 models →
Lyria 3.5Lyria 3 ProElo (Artificial Analysis arena) · 13 models · top 5 coloured · axis does not start at zeroSource: artificialanalysis.ai ↗ · checked 2026-10-23
Open-weights TTS (AA subset)
Swipe the chart sideways for all 12 models →
Magpie-Elo (Artificial Analysis arena) · 12 models · top 5 coloured · axis does not start at zeroSource: artificialanalysis.ai ↗ · checked 2026-10-23
Subset of the AA Speech board: only models AA marks as open weights. Ranks shown are overall AA ranks.
Score vs price
Voice / TTS
Cost per minute and per song Estimate
TTS: 1 min ≈ 750–1,000 characters (Cartesia states 750–800 credits per minute); token-billed: 25 audio tokens/s (Google).
| Model | $ per minute | $ per song | Basis |
|---|---|---|---|
| Eleven v4 | $0.08 | — | vendor ~$/min |
| Eleven v4 Turbo | $0.04 | — | vendor |
| Gemini 3.8 Flash TTS | $0.0135 | — | $9/M × 1,500 tokens |
| Gemini 3.8 Flash-Lite TTS | $0.0067 | — | $4.50/M × 1,500 |
| gpt-4o-mini-tts | $0.015 | — | third-party estimate |
| Inworld TTS-2 | $0.02 | — | $25/M chars × ~800 chars |
| Qwen-Audio-3.1-TTS-Plus | $0.015 | — | $19.3/M × ~800 |
| MiniMax Speech 2.8 HD | $0.08 | — | $100/M × ~800 |
| Lyria 3.5 | — | $0.08 | vendor list |
| Eleven Music | $0.15 | — | vendor list |
| Eleven SFX | $0.12 | — | vendor list |
Open source and ComfyUI
Voice and music models you can run yourself.
| Model | Licence | ComfyUI | Minimum VRAM | Rank |
|---|---|---|---|---|
| Kokoro 82M v1.0 | Apache 2.0 | community nodes (not verified) | Official: 82M params; runs on CPU or small GPUs (size-based inference, no official VRAM figure) huggingface.co ↗ artificialanalysis.ai ↗ | tts · AA TTS #49 (1085) |
| F5-TTS | CC-BY-NC-4.0 (weights) | community nodes | Official: not stated huggingface.co ↗ | tts+cloning |
| Dia 1.6B (Nari Labs) | Apache 2.0 | community nodes | Official: ~10 GB full version (model card) huggingface.co ↗ | dialogue tts |
| Sesame CSM-1B | Apache 2.0 | community | Official: not stated huggingface.co ↗ | conversational speech |
| VibeVoice 1.5B (Microsoft) | MIT | custom nodes (Enemyx-net, wildminder) | Official: not stated Community report: community: 1.5B ~6–8 GB; 7B ~16–20 GB; 4-bit ~7–8 GB huggingface.co ↗ github.com ↗ | podcast / multi-speaker · AA TTS #83 |
| Chatterbox (Resemble AI) | MIT | community | Official: not stated huggingface.co ↗ | tts+cloning, 23 languages · AA TTS #72 |
| Fish Audio S2 Pro | fish-audio-research-license | not verified | Official: not stated huggingface.co ↗ | tts, 83 languages · AA TTS #38 (1116) |
| Breeze TTS 2 (BreezeBlue) | open weights per AA (licence not checked) | — | Official: not stated artificialanalysis.ai ↗ | tts · AA TTS #13 (1222), highest-ranked open-weights TTS |
| Voxtral TTS (Mistral) | open weights per AA (HF repo gated) | — | Official: not stated artificialanalysis.ai ↗ | tts · AA TTS #47 |
| Higgs Audio V3 (Boson AI) | open weights per AA | — | Official: not stated artificialanalysis.ai ↗ | tts · AA TTS #61 |
| ACE-Step 1.5 / v1-3.5B | Apache 2.0 (v1 card; 1.5 not checked) | native templates (docs.comfy.org) | Official: official: under 4 GB with the DiT alone, LM planner off (ComfyUI 1.5 docs); full song in <10 s on an RTX 3090 docs.comfy.org ↗ huggingface.co ↗ | music |
| Stable Audio Open 1.0 | stable-audio-community | native (2.3 GB checkpoint) | Official: not stated Community report: community: ~4–6 GB (unofficial mirror of ComfyUI docs) docs.comfy.org ↗ huggingface.co ↗ | sfx / music loops (47 s) |
| YuE s1 7B | Apache 2.0 | community | Official: not stated Community report: 8 GB with quantized models (community UI noted on card) huggingface.co ↗ | full-song music with vocals |
Sources and gaps
- Artificial Analysis TTS Arena ↗ · 97 models
- AA-Music-Vocal v1.1 ↗ · 13 models
- Coval TTS latency/WER ↗ · board 2026-10-16
- TTS Arena V2 (Hugging Face) ↗ client-rendered; not parsed (gap)
- TTS Arena V2 not parsed
- AA board prices encoded; only some captured
- Hume Octave, Udio, Mureka, MiniMax Music prices and dates missing
- Cartesia and Inworld X handles not found
- Most open-source VRAM figures unofficial or missing (F5, CSM, Chatterbox, Fish)
- No verified brand ad case studies for audio
- Teammate input not gathered
FAQ
Which AI voice model sounds most human?
On the Artificial Analysis speech arena, Eleven v4 Turbo leads (1335), just ahead of Eleven v4 (1329). Checked 2026-10-17.
Which AI music model is best?
Suno v6 leads the AA-Music-Vocal board (1146). Suno allows commercial use only for songs made on a paid plan.
Can I run a voice model locally?
Yes. 13 open models are listed with licence, ComfyUI support and minimum VRAM where known; community figures are labelled.