Explore voices
Hear every voice Voxie speaks with.
Every voice reads the same two lines in its language. Voxie's own voices run on your hardware; the others are providers you plug in with a key.
Click a voice's name to see its ratings. Or rate voices blind, without seeing who made them.
Each clip is the voice as it came from its provider, levelled to the same loudness. Voxie picks the voice per language from your config, and moves to another when one fails. Cartesia, ElevenLabs, Deepgram and others plug in the same way, with your own key.
Voices marked not in Voxie yet are ones we're evaluating: Voxie can't speak with them today. Chatterbox Multilingual is one of them. Each character uses one of Voxie's Kokoro voices as its reference, so it sounds like the same person in every language. On our test GPU it needs 2–2.9 seconds to make each second of speech, too slow for a live call, and when we checked every clip with speech-to-text, 19 of 30 read back at 95% or better; the rest slip a word or add a stray sound between sentences. We show them as they came out.
Indic Parler-TTS (AI4Bharat) speaks Indian languages with the speakers it recommends: 16 of 19 clips read back at 95% or better, the other three at 91%. It's slower still, 3.1–4 seconds per second of speech on the same GPU. One more Tamil speaker we tried, not among its recommended ones, came out as gibberish and is left out.
Chatterbox Multilingual by Resemble AI is MIT; its audio carries Resemble's inaudible watermark. Piper voices under CC BY 4.0: cs_CZ-kasandra (voice and recordings by Ondřej Šimek), sv_SE-alma (NST Swedish dataset, Språkbanken), vi_VN-vais1000 (VAIS-1000 corpus), it_IT-serena (serena-synthetic-it-27h dataset by committa). Kokoro is Apache 2.0; Indic Parler-TTS by AI4Bharat is Apache 2.0. ElevenLabs clips were made with ElevenLabs (elevenlabs.io); OpenAI clips are AI-generated voices from OpenAI's gpt-4o-mini-tts. All credits: NOTICE.