Call Recording Transcription for Asterisk, FreePBX and Vicidial

Upload a call recording and get back a written transcript with timestamps, split into caller and agent when the recording is stereo. It reads the files Asterisk, FreePBX, Issabel and Vicidial write to disk, including raw .gsm, GSM-in-WAV (WAV49), G.711 µ-law and A-law and signed-linear .sln, so you do not have to convert anything first. Download the result as plain text, SRT subtitles or JSON.

Short answer: Choose the recording, leave the language on “Detect automatically”, tick the consent box and press Transcribe. A 3-minute call usually comes back in well under a minute. Transcription runs on srvScripts’ own servers with the open-source Whisper model; recordings are not sent to any third-party AI service.

How it works

1. You choose a recording. The page checks the length in your browser so you know straight away whether it fits your plan.
2. The file is uploaded over HTTPS directly to the transcription server. WordPress only hands out a short-lived upload ticket; it never stores the audio.
3. The server converts the audio to 16 kHz mono (or keeps both channels for stereo calls), skips long silences such as hold music gaps, and runs the speech model.
4. You see the transcript line by line with timestamps, and can copy it or download TXT, SRT or JSON.
5. The recording is deleted from the transcription server as soon as the transcript is ready.

Supported recording formats

Format	Where you find it	Notes
.gsm (raw GSM 06.10)	Asterisk / FreePBX default for older systems, Vicidial	Headerless 8 kHz; read directly, no conversion needed
.WAV / WAV49 (GSM in WAV)	Asterisk format=wav49, many Vicidial installs	Always mono, so caller and agent cannot be separated
.wav (PCM, µ-law, A-law)	FreePBX default, call centre exports	Stereo files can be split into caller and agent
.ulaw, .alaw, .sln, .sln16	Asterisk MixMonitor raw formats	Headerless; the extension tells the server how to read it
MP3, M4A, OGG/Opus, FLAC	Cloud PBXs, 3CX exports, converted files	Any sample rate

If a file does not play anywhere, convert it first with the call recording converter (it runs in your browser) to check it is a real recording.

Limits

	Without an account	Free account	Pro
Recordings	2 a day	3 a day	60 hours a month
Length per file	Up to 3 minutes	Up to 30 minutes	Up to 5 hours
Model	Standard	Standard	Standard or Accurate
Caller / agent split (stereo)	Yes	Yes	Yes
Downloads	TXT, SRT, JSON	TXT, SRT, JSON	TXT, SRT, JSON

How accurate is it on phone audio?

We measured both models on a 2 vCPU lab server without a GPU, using 232 seconds of recorded speech from six speakers (LibriSpeech, CC BY 4.0). We then converted the same audio to the two formats phone systems use most, 8 kHz G.711 µ-law and 8 kHz GSM, and transcribed it again. The figures are word error rate: the share of words that were wrong, missing or added. Lower is better.

Model	Clean audio	8 kHz µ-law (phone line)	8 kHz GSM (Asterisk .gsm)	Speed on 2 vCPU
Standard (Whisper base)	4.8%	5.3%	7.3%	about 6× faster than real time
Accurate (Whisper small)	2.7%	2.2%	3.8%	about 2× faster than real time

In practice that means roughly one word in 25 is wrong on a GSM recording with the accurate model. Real calls are harder than our test audio: background noise, people talking over each other and strong accents raise the error rate. Names, e-mail addresses and account numbers are the words most often misheard, so check those against the recording.

In a stereo test call (one speaker per channel, 8 kHz µ-law) with caller / agent split on, the accurate model scored 0.0% errors on the caller side and 2.1% on the agent side.

Getting caller and agent on separate channels

The transcript can only name “Caller” and “Agent” if they are on different channels of a stereo recording. Asterisk’s MixMonitor mixes both directions into one mono file by default. Its r() and t() options also write the received and transmitted audio to separate files, which you can join into one stereo WAV with sox:

; extensions.conf - record both legs separately as well as the mixed file
exten => _X.,n,MixMonitor(${UNIQUEID}.wav,r(${UNIQUEID}-in.wav)t(${UNIQUEID}-out.wav))

# afterwards: -in = what this channel sent (the caller, when MixMonitor runs on the
# caller's channel), -out = what it heard (the agent). Left = caller, right = agent.
sox -M 1700000000.123-in.wav 1700000000.123-out.wav 1700000000.123-stereo.wav

Upload the -stereo.wav file and tick “Stereo recording”. GSM-in-WAV (WAV49) is always mono, so record in wav if you want the split.

Privacy and consent

• Recordings are processed on servers run by srvScripts. They are not sent to OpenAI, Google or any other AI provider.
• The recording is deleted as soon as the transcript is ready. The transcript is kept for 24 hours so you can download it again, then deleted.
• We do not use your recordings or transcripts to train models.
• Call recordings usually contain personal data. Make sure you are allowed to process them: many countries require callers to be told the call is recorded, and some require every party’s consent.

Frequently asked questions