The microphone needs a secure page.
Open the HTTPS address and accept the self-signed certificate once.
Or, on this Mac, run ssh -L 8770:localhost:8770 agx-ts and use localhost.
Transcript
Say something.
Tap the button or press Space. Words appear while you are still talking and settle when you pause. Your audio is transcribed on the Jetson and goes nowhere else.
Model
The last few models you used stay loaded on the Orin, so switching between them is instant. The first pick of a new one downloads it once.
Voice
Each translation is spoken aloud in its own language, by the Jetson. Pick a voice for every language; choosing one plays a short sample.
Speak translations
The microphone pauses while the device speaks, so it never transcribes itself.
Delivery
For the natural voices: they follow a description of how to speak.
Pace
Pitch
Style
Asking the Jetson which voices it has…
Tuning
Trade a little patience for whole sentences, or the other way round.
Auto-detect chooses from
Used when Language is "Auto · Indian + English". The fewer languages, the fewer mix-ups: tick only the ones your speakers use.
When someone speaks the language the page translates into (English, translating into English), their line is kept as it is by default: shown and spoken in that language. Or send it the other way: into the other language of the conversation (the one heard most lately, Hindi until another is heard), or into a language you pick.
How long a pause settles a line. Shorter settles sooner but can split sentences.
Live words while speaking
Re-transcribes back-to-back while you talk. Off saves GPU.
Settled line
What it's hearing
−70 dB−35 dB0 dB
Top bar: the microphone, with a mark at the room's noise floor. Bottom bar: how sure it is that this is speech — a line starts when it crosses the mark.
Start from
How sure it must be before it starts a line. More eager catches a quiet talker; too eager and a fan becomes a sentence. Watch the bottom bar while you speak.
Lifts a quiet voice before anything listens to it. Measured: 24 dB takes a talker at −55 dBFS from 56% of their speech detected to 79%.
Sound this far below the person speaking counts as silence, so someone across the room stops holding your line open.
Mixes a denoised copy into what the recognisers hear, never the detector. Helps against a fan or nearby talk; costs a few points on clean speech.
Play back what the recogniser heard
Each line can be played from the transcript. Off plays the microphone as it arrived; on plays the same audio after the gain and the clean-up, which is what it actually worked from.
