48× realtime on M4 — transcribes 11 seconds of Russian speech in 230ms. Compared to the CTC version, RNNT offers ~9% lower Word Error Rate (WER) across benchmarks due to the autoregressive joint language modeling loop, with slightly slower sequential decoding.

Original model: ai-sage/GigaAM

Quick Start

pip install mlx safetensors numpy

python

from huggingface_hub import snapshot_download

model_dir = snapshot_download("al-bo/gigaam-v3-rnnt-mlx")

Or use with the inference code from GigaAM MLX:

python

from gigaam_mlx import load_model, load_audio

model = load_model("./gigaam-v3-rnnt-mlx")
text = model.transcribe(load_audio("audio.wav"))
print(text)
# → ничьих не требуя похвал счастлив уж я надеждой сладкой

Architecture

Audio (16kHz) → Log-Mel Spectrogram (64 bins)
             → Conv1d Subsampling (4× stride)
             → 16× Conformer Layers:
                  ├─ FFN₁ (half-step residual)
                  ├─ RoPE Multi-Head Self-Attention (16 heads)
                  ├─ Convolution Module (GLU + depthwise conv)
                  └─ FFN₂ (half-step residual)
             → RNNT Head (Joint + LSTM Decoder)
             → Greedy Decode

Performance (Apple M4)

Metric	Value
Batch (11s audio)	230ms (48× realtime)
Model size	423 MB (fp16)
Parameters	~222M

Files

model.safetensors — weights (fp16, 423 MB)
config.json — model config + vocabulary (34 Russian characters)

Conversion

Converted from PyTorch using convert_gigaam_to_mlx.py. LSTM weights are transformed from PyTorch (weight_ih, weight_hh, bias_ih, bias_hh) to MLX layout (Wx, Wh, bias).

License

MLX conversion code: MIT. Model weights: see ai-sage/GigaAM license.