Dimitrius174/gigaam-v3-onnx-ru

GigaAM-v3 ONNX — Russian Speech Recognition

GigaAM-v3 (222.5M params) exported to ONNX with Int8 quantization. CER ~4-5%, RTF ~0.04 on CPU (25× faster than real time).

Quick Start

pip install onnxruntime numpy torch torchaudio soundfile pandas

python3 recognize.py speech.wav

Result: "В древнем Китае использовали уникальный способ обозначения периодов времени."

Знаки препинания (заглавные буквы, запятые, точки) восстанавливаются автоматически через Silero TE model (скачивается при первом запуске).

Без пунктуации: --no-punct

Model Files

File	Size	Description
model/gigaam_v3_rnnt_encoder.onnx	305 MB	Conformer encoder (Int8)
model/gigaam_v3_rnnt_decoder.onnx	3.2 MB	LSTM decoder
model/gigaam_v3_rnnt_joint.onnx	1.4 MB	Joint network
model/gigaam_v3_rnnt_tokens.txt	195 B	Token vocabulary (33 chars + blank)

Architecture

• Encoder: Conformer 16 layers, 768 dim (222M params)
• Decoder: LSTM 1 layer, 320 dim + Embedding (34→320)
• Joint: 2× Linear(768→320, 320→320) + ReLU + Linear(320→34)
• Vocabulary: 33 Russian letters + space + <blk>

Performance

Metric	Value
Params	222.5M
CER (test set)	~4.9%
RTF (CPU, FP32)	0.04
RTF (CPU, Int8)	0.03
Avg decode (12s audio)	~450 ms (FP32) / ~320 ms (Int8)

Usage

Command Line

# Recognize a WAV file (16kHz, mono)
python3 recognize.py speech.wav

# Benchmark
python3 recognize.py --benchmark

Python API

from recognize import GigaAMRecognizer

asr = GigaAMRecognizer('model')

# From file
text = asr.transcribe('speech.wav')

# From numpy array (16kHz, mono)
import soundfile as sf
audio, sr = sf.read('speech.wav')
text = asr.transcribe_raw(audio)

print(text)  # "эта идея пришла из китая где излюбленным цветком был цвет сливы"

GUI (offline ASR app)

python3 asr_gui.py

Buttons: 🎤 Start / ⏹ Stop. Text is copied to clipboard automatically.

Requirements

• Python 3.10+
• onnxruntime (or onnxruntime-gpu for GPU)
• numpy
• torch, torchaudio (for mel spectrogram)
• soundfile
• pandas (optional, for dataset evaluation)
• Silero punctuation model (auto-downloaded via torch.hub, ~87MB)

License

MIT. The underlying GigaAM-v3 model is from ai-sage/GigaAM-v3.

Citation

@software{gigaam_v3_onnx,
  title = {GigaAM-v3 ONNX: Russian Speech Recognition},
  author = {Sber AI},
  year = {2024},
  url = {https://huggingface.co/ai-sage/GigaAM-v3}
}