packethushRequest access

Hear it

Hear it in your language.

Indian phone calls with a second voice and street noise, at 8 kHz as a phone line carries them. Switch between the original and packethush while it plays. These calls are in Hindi. Comparisons in Kannada, Tamil and other Indian languages are coming soon.

hindi

Outside: a second voice beside the caller, and the street.

second voice
next to the phone, 10 dB quieter than the caller
second voice after packethush
26 dB below the caller, on average
caller level change
0.0 dB
0:00 / 0:20
original

Ma'amमुझेआपकेhotelमेंroombookकरनाथा.यहांहरुईहरुईजीबहनभेजनाहै.Ma'amदोसेतीनलोगहैं.आपमुझेroomकेबारेमेंजानकारीदेसकतेहैं?अबवहएकएकkiloतकहोगयाहै.मैंआपकाhotelमेंrestaurantहैयाहमेंबाहरसेखानाorderकरनापड़ेगा?

packethush

Ma'am,मुझेआपकेhotelमेंroombookकरनाथा.Ma'am,दोसेतीनलोगहैं.आपमुझेroomकेबारेमेंजानकारीदेसकतेहैं?आपकाhotelमेंrestaurantहैयाहमेंबाहरसेखानाorderकरनापड़ेगा?

extra words, struck out: 15 → 0 · counted against what Deepgram nova-3 writes for the caller alone

hindi

The same call at home: a second voice across the room, as loud as the caller.

second voice
across the room (echoey), as loud as the caller
second voice after packethush
13 dB below the caller, on average
caller level change
0.3 dB
0:00 / 0:20
original

Ma'amमुझेआपकेhotelमेंroombookकरनाथा.यहांभेजनाहै.Ma'amहमदोसेतीनलोगहैं.आपमुझेroomकेबारेमेंजानकारीदेसकतेहैं?सबphoneएकएकkiloसबphoneहै.Ma'amआपकाhotelहै.Restaurantहैयाहमेंबाहरसेखानाorderकरनापड़ेगा?

packethush

Ma'am,मुझेआपकेhotelमेंroombookकरनाथा.Ma'am,दोसेतीनलोगहैं.आपमुझेroomकेबारेमेंजानकारीदेसकतेहैं?आपकाhotelहै,restaurantहैयाहमेंबाहरसेखानाorderकरनापड़ेगा?

extra words, struck out: 11 → 0 · counted against what Deepgram nova-3 writes for the caller alone

Voices: IndicVoices by AI4Bharat (CC BY 4.0) · street sounds: Freesound (CC0) · levels measured on each call · more languages and scenes are being recorded

Proof it streams

Change the end of a call. The start stays the same.

We ran one call through packethush twice: once as it was, once with everything after 8 seconds replaced by a different call. A batch model reads the whole call before it answers, so a new ending would change its output from the start. packethush only uses audio that has already arrived.

input, run 1
input, run 2
output, both runs
8.00 s
0 of 64,000 output samples differ before the change87,542 of 99,520 differ after it · run 1 run 2

Hear it on your own calls.

The trial includes a command-line tool and an evaluation kit that process your recordings the same way, 20 ms at a time, on your machine, and measure the result with your speech engine.

Request access