Georgian speech to text, built for telephone calls
We are building a streaming speech recognition model for Georgian phone audio, the 8 kHz narrowband sound of call centers, banks, IVR and voice agents. It runs in your own infrastructure, so audio never leaves your environment.
Status: In development · model in training · early access waitlist open
The problem: Georgian on the phone
Speech recognition has become good for many languages, but Georgian telephone audio is still a hard case. General-purpose models are trained mostly on wideband recordings and high-resource languages. On real Georgian calls they make too many mistakes to be useful.
Narrowband audio
Phone lines carry 8 kHz audio. Everything above 4 kHz is gone, including much of the detail that separates sounds like ს/შ or ც/ჩ.
Real-world channels
Codecs, packet loss, background noise, echo and crosstalk degrade the signal further.
Real conversations
Callers mix Georgian with Russian and English words, and say numbers, names and addresses that matter most to get right.
What we are building
Streaming
Partial transcripts while the caller is still speaking, so voice agents and live tools can react in time.
Telephony first
Fine-tuned on real 8 kHz telephone speech, not adapted as an afterthought from wideband audio.
Private deployment
Runs on-premises or in your own cloud account. Audio and transcripts stay with you.
One focus
Georgian only, telephony only. We are not a general-purpose platform.
Where we are today
Graey is in early R&D. The model is in training on NVIDIA NeMo. There is no public release yet, no customers, and no published accuracy results. We will share evaluation numbers once we have measured them on real, held-out Georgian telephone audio.
Facts at a glance
A plain summary you can quote.
- What
- Streaming speech-to-text (automatic speech recognition, ASR) for Georgian
- Language
- Georgian (ქართული)
- Audio
- 8 kHz narrowband telephone audio
- Mode
- Low-latency streaming with partial and final transcripts
- Deployment
- On-premises or in the customer’s own cloud; audio stays in the customer’s environment
- Technology
- NVIDIA NeMo, cache-aware streaming FastConformer, fine-tuned on telephone speech
- Stage
- In development. Model in training. No public release, no published benchmarks yet. Early access waitlist open.
- Company
- Graey, Tbilisi, Georgia. Founded 2020.
- Contact
- [email protected]
Working with Georgian phone audio?
Join the early access waitlist. When the model is ready for testing, we plan to start with a small group of call centers, banks and voice-agent teams on their own audio.