asterisk

Uttera + Asterisk

Versión en castellano

Two ways in, and the second one needs no change to your dialplan at all.

agi/           speak in a live call, and hear the caller
recordings/    transcribe the calls your PBX is already recording
panel/         a web view of the call log, with transcript and intelligence

recordings/ — start here

Your PBX is already writing .wav files somewhere. Point the script at that folder and last week's calls are transcribed, without touching Asterisk. This is the code we run in production. See recordings/.

panel/ — see the calls

A small, read-only web page over your call log. It joins the CDR with the transcripts and intelligence recordings/ already leaves next to each .wav, and shows one row per call with its transcript, sentiment, speakers and caller profile. For the calls that matter, one click asks api.uttera.ai for a narrative summary or a signed PDF. Stdlib only, four languages, put it behind your HTTPS proxy. See panel/.

agi/ — talking inside the call

Two AGI scripts: one speaks a text in the call, the other listens to the caller and leaves what they said in a dialplan variable.

exten => 101,1,Answer()
 same => n,AGI(uttera-decir.agi,"How can I help you?")
 same => n,AGI(uttera-oir.agi,8,en)
 same => n,NoOp(The caller said: ${UTTERA_TEXTO})

Installation

cp agi/uttera_agi.py /var/lib/asterisk/agi-bin/
chmod +x /var/lib/asterisk/agi-bin/uttera_agi.py
ln -s uttera_agi.py /var/lib/asterisk/agi-bin/uttera-decir.agi
ln -s uttera_agi.py /var/lib/asterisk/agi-bin/uttera-oir.agi
apt install sox

The key goes in the service environment, never in the dialplan — a key written there ends up in your configuration backup, in version control and in the output of dialplan show:

systemctl edit asterisk
[Service]
Environment=UTTERA_API_KEY=sk-echo-...

What costs a day to find out

All of it is handled in the code. It is written down because it takes a day to discover and five minutes to read.

The audio conversion is not optional. Uttera generates at 24 kHz and a phone channel runs at 8 kHz. Hand Asterisk the WAV as it comes and it plays it at its own rate: the voice comes out fast and high-pitched. sox converts to .sln, raw PCM at 8 kHz, which is exactly what the channel wants.

Add silence at the end. Without a few tenths of tail (pad 0 0.4), Asterisk clips the last syllable when it closes the file.

STREAM FILE takes no extension. Give it the base name and it picks the format it finds.

Whisper does not stay quiet when given silence. It does not return an empty string — it invents a sentence, usually something like "Thanks for watching", because that is what its training data is full of. The AGI drops recordings under 2 KB before sending them: it saves you paying for a hallucination and, worse, acting on it.

The quoting in SET VARIABLE and VERBOSE. The AGI parser splits on spaces: unquoted text arrives truncated and the variable keeps only the first word. A quote inside the text derails the rest of the line, so they are replaced before sending.

Telephony timeouts are not API timeouts. Uttera holds the connection for up to two hours for long jobs, but here there is a person with a phone to their ear: these scripts give up after 30 seconds. If Uttera has not answered by then the call is already ruined, and the thing to do is say so, not keep waiting.