Voice Bake-off

25 text-to-speech voices, from free on-device to commercial APIs, compared for a spoken status-update assistant. Each card shows how long it took to generate and what 30 seconds of speech costs.

Test script · 84 words · 490 characters · ~30s

Morning! Quick status check: everything's up and running. All five sites responded normally over the last hour, and last night's builds all passed. You've got three new support tickets. Two are billing questions, and one is a bug report about the export button, which is probably worth a look today. Signups are up about twelve percent over last week. Nothing urgent on the security side. If you only do one thing this morning, answer that bug report before it turns into a refund request.

Commercial vendor cards (marked “vendor demo”) play each vendor's own published sample instead — their cost is calculated for this 490-character script.

How these were measured

First audio — time until the first playable audio. macOS and Kokoro: time to render the opening sentence (how a streaming assistant would use them). OpenAI TTS: time to first byte of the HTTP response. Realtime: time from connect to first audio chunk.
Full render — time to generate the entire ~30s clip in one go.
Cost / 30s — for the 490-character script. Realtime costs are actual billed usage from the API response; gpt-4o-mini-tts is estimated at ~20 audio tokens/sec (the rate observed from Realtime) × $12/1M; vendor costs use each vendor's published pay-as-you-go or cheapest paid-tier rate as of 2026-10-06.
Local runs: Apple Silicon Mac. macOS via say -r 185. Kokoro-82M via mlx-audio 0.5.8 (Kokoro-82M-bf16), timed after a warm-up — the first call after model load takes ~5s.
Vendor demos are linked to their sources and remain the property of their respective owners. Vendor timings weren't measured; where a latency is shown it's the vendor's own claim.