Speechify Simba 3.5 tops Controlled Voice Arena at lowest price

Speechify's new text-to-speech model, Simba 3.5, shares first place on Artificial Analysis's Controlled Voice Arena, the board that measures how well a model speaks in a voice it is given, and it is the cheapest model on Artificial Analysis's price chart. Artificial Analysis ranks speech models by listener votes in blind comparisons of two clips, and it runs two versions of that arena. On Controlled Voice every model clones the same set of reference voices, so the vote judges the speaking rather than the voice. On Provider Voice each model uses voices its own provider offers. According to Artificial Analysis, Simba 3.5 ranks first on Controlled Voice, statistically level with HeyGen Voice, which means the gap between the two ratings is too small for the votes to separate them, and ahead of Qwen-Audio-3.1-TTS-Plus from Qwen. In late October, Artificial Analysis had placed Speechify's earlier model Simba 3.2 outside the top ten on the same board.

The lead does not carry over to the other board. With providers' own voices, Simba 3.5 ranks fifth, level with Cartesia's Sonic 3.6 and behind models including Qwen-Audio-3.1-TTS-Plus and ElevenLabs' Eleven v4 Turbo. It ranks seventh on Pronunciation Robustness, Artificial Analysis's test of saying difficult text correctly. Its clearest advantage is price. Artificial Analysis lists it at $6.58 per million characters and measures it at 139.6 characters per second, against $40 and 105.3 characters per second for Eleven v4 Turbo; the post makes that speed comparison only with Eleven v4 Turbo, not with the whole field. Simba 3.2 was already listed at $6.6, so on these results Speechify has moved to the top of the cloned-voice board at about the same price, without becoming the leader when each provider picks its own voices. Artificial Analysis's leaderboard pages were not read, so Simba 3.5's ratings and uncertainty ranges are not covered here.