Features
Everything packethush does for your agent.
One library, in front of your VAD and speech recognition.
01
Voice isolation
Primary speaker extraction: keeps the caller's voice and removes every other voice near the phone.
- Your agent answers only what the caller said, not the room.
- No enrolment, no voice samples
- Finds the caller from the call itself
- Before your VAD and speech recognition
- 64%fewer words the caller never said (1,152 Hindi test calls)
- 0.709 → 0.485word errors, same calls, Deepgram nova-3
02
Noise cancellation
Noise suppression in the same pass: traffic, fans, crowds, music and line hiss.
- Less noise reaches your agent's speech recognition and VAD.
- No extra step, no extra delay
- Nothing lost when the caller is alone
- Noise between words 38 dB lower
- 0.272 → 0.271word errors, caller alone
- 0.513 → 0.469word errors, loud street noise (5 dB SNR)
03
Voice activity and turn-taking
When the caller starts, stops and finishes a turn, from the caller's voice alone or from every speaker.
- Your agent takes turns with the caller, not with the other voices in the room.
- Fewer false triggers than Silero VAD, from other voices and from background noise
- Fewer of the caller's turns missed than with Silero VAD
- Start, stop and end-of-turn events, timed by your settings
- From the same pass that cleans the audio: no second model
Against Silero VAD on our Hindi phone-call tests, with the same settings. Results depend on the settings you use.
Silero-style VADComing soon
- Outputs
- speech probability every 20 ms · speech start · speech end · end of turn · speech segments with times
- Settings
- threshold · exit threshold · minimum speech · minimum silence · speech padding · maximum speech length
- For
- the caller alone, or every speaker · 8 or 16 kHz
- 6settings, as in Silero VAD: you set when a turn starts and ends
- 20 msa fresh speech probability with every packet
04 · Coming soon
Audio insight
A risk score for the audio of every call, live and after the call, with its causes.
- During the call: your agent sees the risk rise and can ask the caller to move somewhere quieter, read numbers back, or hand over to a person.
- After the call: you find the failed calls, and why they failed, without listening to them.
- Without extra latency: scored in the same pass that cleans the audio
- Compare carriers, regions and campaigns
What it measuresComing soon
- Risk score
- 0 to 1: how likely the audio is to break speech recognition, VAD and turn-taking
- Causes
- background noise · other voices · echo (caller far from the phone) · packet loss · codec damage · caller loudness
- When
- every few seconds during the call, and per call afterwards: average, worst moments, share of the call at risk
- 0 to 1risk score, live and per call
- 6causes, each scored 0 to 1
05
Every Indian language
Works on the voice, not the words. No language setting.
- One setup for every caller you serve, in any language mix.
- Every Indian language, and English
- Calls that mix languages
- One model for all of them
- 1model for every language
- Hindithe language of our published results
06
Built for phone lines
μ-law, A-law or PCM in, the same format and length out.
- No conversion and no extra hop between the phone network and your agent.
- PCM as 16-bit or float
- 8 or 16 kHz, in 20 ms packets
- Fits inside a live call
- +12 msadded delay
- 3formats: μ-law, A-law and PCM
07
SDKs and integrations
Python, Go, Rust and C, a FreeSWITCH module and a WebSocket bridge.
- It fits the stack your agent already runs on.
- The same samples from every package
- One session per call leg
- LiveKit and Pipecat plugins: coming soon
- 3lines of code
- 4languages, one engine
Hear every feature on your own calls.
Your recordings, your speech engine: hear and measure the difference.