Models for VRChat Admin Tools

Files the VRChat Admin Tools translator downloads for on-device Russian voice input. Nothing here is used on a server: the program runs the model on the player's computer with ONNX Runtime Web.

gigaam-v3-e2e-ctc

GigaAM-v3 e2e_ctc by the GigaChat Team (SberDevices), MIT license (see LICENSE), converted to ONNX fp16 without retraining.

  • Source: ai-sage/GigaAM-v3, revision e2e_ctc, commit cec030b4c4f35d928e4a9044a3bdb29ebd499fac.
  • Conversion: the model's own CTC ONNX export (opset 17), then onnxconverter-common 1.16.0 fp16 with fp32 inputs and outputs; 33 Cast nodes retyped to fp16 so ONNX Runtime accepts the graph.
  • Inputs: features float32 [1, 64, T] (log-mel: 16 kHz, n_fft 320, hop 160, Hann, 64 HTK mels, natural log clamped to 1e-9), feature_lengths int64 [1]. Output: log_probs float32 [1, T', 257], class 256 is the CTC blank.
  • vocab.json: SentencePiece pieces by id (256).
  • Check: on 12 test recordings the ONNX fp16 text equals the original PyTorch model's text on every recording.
File Size SHA-256
gigaam-v3-e2e-ctc/v3_e2e_ctc.fp16.onnx 443438383 9343dcfc4cb39e0564abbbc7120d1c2684a3194f52b400ebd03ce49a006a8918
gigaam-v3-e2e-ctc/vocab.json 1852 9c97685c7510b3f1601362b9521989cf559af0121c85de3c7ab57f5b4c967c4c
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for Rose337/VRChatAdminToolsModels

Quantized
(38)
this model