Internal · TTS test bench

TTS Lab

One sentence, every model. Tune each voice individually and compare by ear — costs shown per clip as you type.

Access code

Sentence to speak

Live

Kokoro-82M via OpenRouter

Open-source budget voice. Chinese is an added language — fine, but not native prosody. No tuning: what you hear is what you get.

$0.62 / 1M chars ≈ — Pay-per-use · no subscription

Voice

This is everything OpenRouter exposes for Kokoro — voice choice only (it even hangs on a speed param). Output fixed to mp3.

Live · needs fal balance

Qwen3-TTS (open-source) via fal.ai

The native-Mandarin family, open-sourced by Alibaba and hosted on fal. Takes an optional free-text prompt plus full sampling controls. Voice cloning possible later via fal's clone endpoint.

$90 / 1M chars (1.7B) · $70 (0.6B) ≈ — Pay-per-use via fal · no subscription · top up at fal.ai/dashboard/billing

Model Voice Language Prompt — exactly this text is sent (edit or delete freely; empty = none sent)

Sampling & clone settings (their defaults shown)

Temperature 0.90 higher = more varied delivery Top-p 1.00 Repetition penalty 1.05 Top-k Max new tokens Subtalker sampling their default: on

Voice-clone embedding URL (optional) Clone reference text (optional)

Live

ElevenLabs Multilingual v2 direct API

Premium voices, very natural — but English-first doing Chinese, so listen for accent. Tune with the dials, or paste any voice ID from the voice library to try it.

1 credit / char (≈ $150–220 / 1M) ≈ — Free 10k credits/mo (testing) · Starter $5/mo 30k · Creator $22/mo 100k

Model Voice Custom voice ID (optional) Language code Seed (optional) Text normalization

Stability 0.50 low = expressive · high = consistent Style 0.00 higher = more exaggerated Similarity 0.75 adherence to the original voice Speed 0.90 0.9 ≈ teacher pace Speaker boost clarity boost (their default: on)

v3 only honours stability (its API ignores the other dials — they apply to the v2-family models). v3 also reads inline tags in the sentence, e.g. [slowly] [warmly]. Language code applies to Turbo/Flash/v3 only. Not exposed (plumbing, not voice): output format (fixed mp3), streaming, previous/next-text continuity, pronunciation dictionaries.

Other options — can wire in on request

Alibaba's commercial Qwen3-TTS-Flash — the tuned flagship of the family (~$12 / 1M, free-text instructions). Blocked for now: Alibaba Cloud signup wouldn't go through; the open-source version above is its sibling.
OpenAI gpt-4o-mini-tts — takes free-text voice instructions; ~$12 / 1M chars, pay-per-use. Needs a direct OpenAI key (OpenRouter doesn't host it).
CosyVoice 2 (Alibaba, open model) — native Mandarin, very strong tones; ~$1–3 / 1M via Replicate or Fal, pay-per-use.
MiniMax Speech-02 — Chinese provider, excellent Mandarin; pay-per-use API.

If Kokoro / Qwen / ElevenLabs don't satisfy, say the word and any of these gets a card here.