GigaAM v3 E2E RNNT (int8, sherpa-onnx format)

Mirror for Anti-Vocale on-device Russian ASR.

Base model: GigaAM v3 by Sber (Salute Developers), MIT license
Export: end-to-end RNNT transducer in sherpa-onnx nemo_transducer format, encoder quantized to int8
Source of these files: amidexe/govorun-lite release model-gigaam-v3 (identical bytes; upstream credits the official GigaAM v3 release)
Files: encoder (int8, ~304 MB), decoder, joint, tokens.txt (1024 tokens, native punctuation and capitalization)
Runtime: sherpa-onnx OfflineRecognizer, 16 kHz, greedy search

SHA-256

781971998e6a355d6a714f6932a30eab295e7ba0d14fd7e0f78c83b87e811860  gigaam_v3_e2e_rnnt_decoder.onnx
2cac62d0c270bd128f898f2be1a2d34780d524a6e9483888ebac7b00f97410f1  gigaam_v3_e2e_rnnt_encoder_int8.onnx
602ff7017a93311aad34df1437c8d7f49911353c13d6eae7a6ee7b041339465c  gigaam_v3_e2e_rnnt_joint.onnx
7ddf22514c42c531358182c81446a8159771e9921019f09ae743ea622d40221d  gigaam_v3_e2e_rnnt_tokens.txt

Downloads last month: -; Downloads are not tracked for this model. How to track