MEGAVOX V2
Listen on this machine.
Twelve languages on the A10G. Short clips return a wav file. Long passages stream over a WebSocket and start playing as the first frames arrive.
Numbers, dates, and amounts are always spoken. 12-10-24 becomes “twelve October two thousand twenty four”, and ₹2,500 becomes “two thousand five hundred rupees”. Emotion tags stay as written.
Emotion tags sit in the sentence, in square brackets, next to the words they should color. The server still reads numbers and dates aloud and leaves the tag itself untouched. They are best effort: [laughs] [sighs] [whispers] [sad] [excited] [clears throat]. A tag changes the delivery from that word on, so [whispers] in the middle of a Siddhant line sounds like a second voice. The long sample stays one voice; add a tag yourself when you want the shift.
This server
Live figures come from GET /health. A frame is 80 ms of audio. Tick time is how long the GPU spends producing one frame for the current batch. Time to first audio is measured on the stream itself.
A passage longer than about twenty seconds is split at a sentence boundary and each piece is spoken as a new take, with a 150 ms gap, on one WebSocket call to /v1/tts/stream. The voice can shift at that join. The limit is 4,000 characters.