DmitrySharonov/GigaAM-Multilingual-ONNX

GigaAM Multilingual ONNX

FP32 ONNX exports of all four revisions of ai-sage/GigaAM-Multilingual: ssl, ctc, large_ssl, and large_ctc.

The original model and its code are licensed under the MIT License. These files were exported with the source model's built-in to_onnx() implementation.

Files

Directory	Source revision	Purpose	Parameters	ONNX file
ctc/	ctc	Ready-to-use CTC ASR	220M	885 MB
ssl/	ssl	Self-supervised speech encoder	220M	885 MB
large_ctc/	large_ctc	Ready-to-use CTC ASR	600M	2.34 GB
large_ssl/	large_ssl	Self-supervised speech encoder	600M	2.34 GB

Every directory contains the ONNX graph, the source model YAML configuration, and validation.json with versions and PyTorch-versus-ONNX numerical results.

ONNX interface

The ONNX graphs contain the neural encoder and, for CTC variants, the CTC head. Audio loading and log-mel feature extraction remain outside the graph, matching the upstream ONNX export design.

Inputs:

• features: float32 [batch, 64, frames] log-mel features;
• feature_lengths: int64 [batch] feature lengths.

Outputs:

• CTC: log_probs float32 [batch, encoded_frames, 71] and encoded_lengths;
• SSL: encoded float32 [batch, hidden_size, encoded_frames] and encoded_len.

Feature extraction settings are included in each YAML file. All variants use 16 kHz mono audio, 64 mel bins, n_fft=320, win_length=320, hop_length=160, and center=false.

Validation

The graphs were checked with onnx.checker and executed with ONNX Runtime on a real 16 kHz Russian speech sample. Outputs were compared against the matching PyTorch source revision. CTC validation additionally requires identical greedy transcriptions.

Revision	Max absolute difference	Mean absolute difference	CTC text match
ctc	3.748e-4	7.626e-6	exact
ssl	6.437e-6	7.488e-7	n/a
large_ctc	2.317e-4	7.788e-6	exact
large_ssl	1.159e-5	8.249e-7	n/a

Validation environment: PyTorch 2.10.0, ONNX 1.22.0, and ONNX Runtime 1.27.0.

Source revisions

• ctc: 2f8a57144e6ec3adfd32fe0484d9ea9913305bc8
• ssl: ac7c6db08133f83478451a659f8470ee8ab47a2d
• large_ctc: 3905cd51c3ed4e88c8edf33f3302969ba480a327
• large_ssl: c459e9d21c1c61a0d4b83fc37c1b5cadd1657506

See the upstream model card for training data, benchmarks, intended languages, limitations, and citation.