Subscribe to GEN
Login to GEN
Speech recognition was adequate for decades and never better than that. Modern models changed it: they cope with difficult accents, attribute speech to each speaker, and work in around 30 languages. GEN run those models locally, on our own hardware or on yours, so recordings are turned into searchable text without ever passing through a public cloud service.
It works the other way too. Text becomes natural speech in several voices and languages, which is how a phone system can greet a caller by name and read them something useful rather than a menu.
We automate the transcription of call recordings and other audio, and make the results available in the formats you need: plain text, a version attributed to each speaker, and a structured file with the full analysis for other systems to read. Transcripts can be summarised, so the point of a long call is clear at a glance.
Once audio is text, it can be checked. We analyse transcripts against required scripts and regulations, and produce exception reports for calls that mention, or fail to mention, specific words or phrases, so compliance is measured across every call rather than a sample.
Online services will do some of this at a reasonable price, but call recordings are among the most sensitive data a business holds. Running the models locally removes that risk entirely.
Around 30 languages are transcribed reliably. The models can attempt more, but accuracy falls away, particularly with tonal languages and those less widely spoken, so we quote only the languages we can stand behind.
Text becomes natural speech in several voices and languages, generated locally like everything else. We are often asked to program telephone systems with it, so callers can serve themselves by voice: a caller recognised from their number in your CRM is greeted personally and offered the options that apply to them, rather than a generic menu.
Combined with speech to text and a language model, the caller can simply ask. Their question becomes text, the model answers it, and the answer is spoken back to them, all within your own systems.
"Good morning Margaret. You have one parcel due for delivery. To track it, press 1. To reschedule it, press 2. To cancel it, press 3."
Margaret presses 1. "Margaret, our driver Dave has your parcel and should deliver it between 2 and 3pm. If that suits you, press 1. To reschedule, press 2."
For FreePBX, GTranscribe is our ready-made transcription module, with search across every transcribed call from inside FreePBX. Transcribed and summarised voicemail, AI call routing and the rest of our work on Asterisk are under telephony development. The same transcription is available to developers through our Arkane API.
No. Transcription runs on local models, on GEN's own hardware or installed on your premises, so call recordings and other sensitive audio never pass through a public speech service.
Around 30 reliably. The models can attempt more, but accuracy falls away, particularly with tonal languages and less widely spoken ones, so we quote the languages we can stand behind rather than the longest possible list.
Yes. Transcripts can be attributed to each speaker, as well as produced as plain text and as a structured file with the full analysis for other systems to use.
Yes. Transcripts can be analysed against required scripts and regulations, with exception reports for calls that mention, or fail to mention, specific words and phrases.
Yes. GTranscribe is our transcription module for FreePBX, with search across every transcribed call from inside FreePBX. Our telephony development page covers the wider options, including transcribed voicemail and AI call routing.