GigaAM v3 RNN-T — CoreML
On-device Russian automatic speech recognition. This repository packages the
GigaAM v3 end-to-end RNN-T model
compiled to Apple CoreML (.mlmodelc) for fast, fully offline inference on
iOS and macOS (Apple Neural Engine / GPU / CPU).
The conversion changes only the serialization format (PyTorch → CoreML) and applies int8 weight quantization; the architecture and weights are unchanged.
Files
| File | Purpose |
|---|---|
Encoder.mlmodelc |
Conformer encoder, traced at a fixed 30 s window (mel T = 3000 → encoder T = 750), int8 weight-quantized. |
Predictor.mlmodelc |
RNN-T prediction network (single-layer LSTM). |
JointDecision.mlmodelc |
Fused joint network + argmax — emits token_id directly, so a greedy decode loop never has to materialize the full [1, 1, 1, vocab] fp16 logits per step. |
vocab.txt |
SentencePiece vocabulary. |
Audio input is 16 kHz mono. The encoder is traced at a fixed 30 s shape, so shorter clips are zero-padded and longer audio must be chunked to ≤ 30 s before inference.
Usage
These are plain compiled CoreML models — load each with MLModel (or
coremltools). The pipeline is: log-mel → Encoder → greedy RNN-T loop over
Predictor + JointDecision, detokenized with vocab.txt.
Attribution & license
This is a format conversion of salute-developers/GigaAM (GigaAM v3). All model credit belongs to the original authors. Please refer to the upstream repository for the authoritative model license and terms of use.