1234 Elo · $27.60 / 1M chars
TEXT TO SPEECH / JULY 2026
Which voice
earns the mic?
Compare what listeners prefer, what APIs charge, and which open-weight voices can live on your Mac.
Current readQwen leads; Simba nearly ties for 64% less.
1230 Elo · $10 / 1M chars
1120 Elo · 83 languages
1.2 GB est. · Apache 2.0
THE SPEECH ARENA
Human preference, priced.
Open weights Hosted only Pareto frontier
Human preference Elo ↑
800
900
1000
1100
1200
$0.50
$1
$2
$4
$8
$16
$32
$64
$128
API price per 1M characters → log scale
HOSTED MODEL
Simba 3.2
Speechify · 8 benchmark voices
- Preference
- 1230 Elo
- API price
- $10.0 / 1M
- Frontier
- Yes
ON YOUR MAC
Private voices, sized.
Weight size is only the start. These conservative working-set estimates leave room for the audio codec, vocoder, runtime, and macOS.
5 of 7 models fit comfortably
Model / best useArena EloEst. memoryLanguage / cloningLicenseFit
Kokoro 82MHexgrad · 82M · Small apps, narration, CPU-friendly experiments1057independent1.2 GB8 lang · fixed voicesApache 2.0ComfortableMagpie MultilingualNVIDIA · 364M · Multilingual speech with a compact footprint1060independent2.1 GB12 lang · fixed voicesNVIDIA open modelComfortableGepard 1.0NineNineSix · 556M · Streaming and zero-shot cloning—not ranked2.7 GB4 lang · clonesApache 2.0ComfortableQwen3-TTS 0.6B BaseAlibaba · 0.9B* · Instruction control and 3-second voice cloning—not ranked3.4 GB10 lang · clonesApache 2.0ComfortableVoxtral 4B TTSMistral · 4B · Expressive voice agents and rapid adaptation1075independent10.5 GB9 lang · clonesCC BY-NC 4.0ComfortableHiggs TTS 3Boson AI · ~4B · Expressive, controlled, multilingual dialogue—not ranked12.0 GB100+ lang · clonesResearch / creator24 GB+Fish Audio S2 ProFish Audio · ~5B · Highest-ranked open weights in this snapshot1120independent13.5 GB83 lang · clonesResearch / non-commercial24 GB+* Qwen markets this checkpoint as 0.6B; its Hugging Face safetensors report a 0.9B total model size. Mac estimates are planning numbers, not measured Apple-silicon benchmarks.
BENCHMARK IT PROPERLY
One voice. Four lenses.
Naturalness
Blind A/B listening with the same prompts. Elo is the clearest public signal here.
Accuracy
Transcribe generated audio and report WER or CER by language. Lower is better.
Latency
Measure time-to-first-audio, real-time factor, and long-form throughput on the exact hardware.
Identity
For cloning, measure speaker similarity and test consent, watermarking, and abuse safeguards.