sentence streaming
It speaks while it is still writing
Text is held only until a sentence ends, or 100 ms of quiet, then it is spoken. You hear the first sentence while the model is still writing the next ones. Bars show order, not measured time.
voice / talk to any of 39 models
Ask out loud and the model of your choice answers in a natural voice, grounded in Prince Singh's real work. It starts speaking at the first full sentence instead of waiting for the whole reply.
you said
Voice is saying
demo replay · English (en-US) · 2 spoken questions per model every 12 hours · audio is never stored
01 / signal path
Audio goes up and speech comes back on the same WebSocket. Nothing is typed, nothing is recorded to disk, and the transcript is the only thing kept.
01Mic48 kHz → 16 kHz
Echo cancellation, noise suppression, then 16-bit mono PCM pushed as you speak.
02Socketone WebSocket
Opened straight to the API; the JWT is checked before the socket is accepted.
03Speech to textAzure · en-US
Continuous recognition hands over each finished phrase.
04Noise filter< 3 chars · 9 patterns
“hmm”, “uh”, stray dots never reach a model or your quota.
05Memoryclose repeat
A question it has answered before is spoken straight from memory.
06Retrieval5 chunks
The same hybrid pipeline as Arx: dense + BM25, fused and graded.
07Model5000-token cap
Your chosen model writes a short answer meant for the ear.
08Sentence buffer100 ms
Text waits for a full stop (or a short pause), then goes to speech.
09Text to speechAndrewNeural
Azure neural voice, casual style, +15% rate, MP3 at 24 kHz.
10Playbackframe by frame
MP3 frames play in order, so you hear sentence one while two is written.
02 / built for the ear
sentence streaming
Text is held only until a sentence ends, or 100 ms of quiet, then it is spoken. You hear the first sentence while the model is still writing the next ones. Bars show order, not measured time.
noise filter
Fillers and stray sounds are dropped before they reach a model or count against your questions.
memory
A question close to one Voice has answered before is spoken straight from memory: no model call, and the answer starts almost immediately. The turn is still saved to your history.
privacy
Sound is streamed for recognition and discarded. Only the transcript of the turn is saved, with the model that answered.
no dead air
If the model fails mid-answer, Voice says a short apology instead of falling silent. Over your limit, it tells you out loud and shows the countdown.
03 / models
Any of the 39 live models can answer out loud, from GPT and Claude to Gemini, Grok and DeepSeek. The model is chosen when the conversation starts and locked while it runs.
From your microphone to a spoken answer on one socket. Click any step to see what runs there and where it lives in the code.
04 / your turn
Ask about Prince's work, his systems or how any of them were built, and hear the answer from the model you pick.
Start talking