Gurbani kirtan speech model (8-bit ONNX)
Used by the on-device live kirtan captions of the Gurbani app (https://gurbani-reader.netlify.app). Runs in the browser with onnxruntime-web; no audio leaves the device.
model.onnx.part0 + model.onnx.part1 joined = one ONNX file (187 MB, sha256
908fb8736bf43243d35d90c180789e7990c959b8e7502b5a8e96d1a3540cb381). tokens.txt is the model's own token list.
Source: surindersinghssj/indicconformer-pa-v3-kirtan
by Surinder Singh, commit aab3be661f94f97ec0d74a587eb1ca01780e75f0, file onnx-pa-only/fp32/indicconformer-pa-ctc.onnx
(a fine-tune of AI4Bharat's indicconformer_stt_pa_hybrid_ctc_rnnt_large). License: Apache-2.0
(https://www.apache.org/licenses/LICENSE-2.0).
Our change: 8-bit weights, per channel (onnxruntime 1.31):
from onnxruntime.quantization import quantize_dynamic, QuantType
quantize_dynamic("indicconformer-pa-ctc.onnx", "model.onnx", weight_type=QuantType.QInt8,
per_channel=True, op_types_to_quantize=["MatMul", "Gemm"])
split -b 94000000 -d -a 1 model.onnx model.onnx.part