subthx/parakeet-v3-onnx

Parakeet TDT 0.6B v3 (redux, ultra) for subthx: ONNX

Parakeet TDT 0.6B v3 (25 European languages) as ONNX for subthx: redux and ultra. The exports are Olicorne's (Olicorne/parakeet-tdt-0.6b-v3-redux-onnx and -ultra-onnx, built on moondream's parakeet-redux and parakeet-ultra); the base model is NVIDIA's parakeet-tdt-0.6b-v3. The upstream cards have no license. subthx adds a 4-bit encoder for GPUs (tools/repack_redux.py) and slimmed fp16 decoders (tools/slim_parakeet_v3.py). Copy the repo's contents into the models folder of subthx.

Files

file	precision / quantization
parakeet-redux/config.json	-
parakeet-redux/vocab.txt	-
parakeet-redux/decoder_joint-model.onnx	fp32 decoder (CPU); weights in the .data file
parakeet-redux/decoder_joint-model.onnx.data	fp32 (as exported)
parakeet-redux/decoder_joint-model.fp16.onnx	fp16 decoder (GPU); subthx slimmed
parakeet-redux/encoder-model.w4.fp16.onnx	4-bit weights, fp16 activations (GPU); subthx repack of the redux encoder
parakeet-redux/encoder-model.w2a8.onnx	2-bit weights, int8 activations (CPU)
parakeet-ultra/config.json	-
parakeet-ultra/vocab.txt	-
parakeet-ultra/decoder_joint-model.onnx	fp32 decoder (CPU); weights in the .data file
parakeet-ultra/decoder_joint-model.onnx.data	fp32 (as exported)
parakeet-ultra/decoder_joint-model.fp16.onnx	fp16 decoder (GPU); subthx slimmed
parakeet-ultra/encoder-model.fp16.onnx	fp16 (GPU)
parakeet-ultra/encoder-model.int8.onnx	int8 (CPU)
nemo128_fbanks.npy	-

Sizes and sha256: manifest.json.