Georgian ASR for narrowband telephone audio
How we approach Georgian speech recognition on the phone: the signal, the model, the data and where it runs.
Why 8 kHz is hard
Telephone networks sample speech at 8 kHz, which keeps frequencies only up to 4 kHz. Most speech models are trained on wideband recordings that go to 8 kHz and beyond, so on phone audio they lose information they learned to rely on.
- Missing high frequencies: fricatives and affricates such as ს, შ, ც, ჩ, ძ, ჯ lose much of their distinguishing energy.
- Codec artifacts and packet loss from VoIP and mobile networks.
- Noise, echo, crosstalk and overlapping speakers.
- Spontaneous speech: hesitations, interruptions, numbers, names, addresses, and Russian or English words inside Georgian sentences.
- Very little public Georgian telephone speech to learn from.
Streaming architecture
We build on NVIDIA NeMo and its cache-aware streaming FastConformer. FastConformer is an efficient variant of the Conformer speech encoder. The cache-aware streaming version processes audio in fixed-size chunks and reuses cached activations from earlier chunks instead of recomputing them.
This keeps latency low and predictable, and the model behaves the same way in training and in live use.
- Chunked inference with limited look-ahead: text appears while the speaker is still talking.
- Chunk size and look-ahead set the trade-off between latency and accuracy.
- Partial hypotheses for live use, final hypotheses for records and analytics.
Data approach
Telephone speech
We fine-tune on real telephone speech, because that is the audio the model will hear in production.
Consented recordings only
We only use recordings we have the right to use, collected with the consent of the people recorded. We do not scrape or buy recordings of private calls.
Separate evaluation
Evaluation sets come from real calls and are kept separate from training data.
Privacy by deployment
Graey runs inside your environment, on-premises or in your own cloud account. Audio and transcripts stay there; nothing is sent to us. We do not use customer audio for training unless the customer explicitly agrees to it in writing.
Accuracy and benchmarks
We have not published accuracy results yet. When we do, we will report word error rate (WER) on real Georgian telephone audio, describe the test set and how it was collected. Until then, any number would be a guess.