You can deploy this model and start earning money today!

runanywhere/
sherpa-onnx-nemo-parakeet-ctc-1.1b-int8

Name: sherpa-onnx-nemo-parakeet-ctc-1.1b-int8
Availability: InStock

sherpa-onnx

onnx

automatic-speech-recognition

nemo

ctc

FastConformer

nvidia/parakeet-ctc-1.1b

quantized

cc-by-4.0

Views

No views yet

sherpa-onnx-nemo-parakeet-ctc-1.1b-int8

A Sherpa-ONNX-loadable INT8 export of NVIDIA's parakeet-ctc-1.1b FastConformer CTC model.

This repository exists for one reason: the upstream INT8 export is missing three ONNX metadata_props entries that Sherpa-ONNX reads to configure its NeMo CTC decoder, so it cannot be loaded as published. This copy carries them, and is otherwise byte-identical to upstream.

Provenance

Derived from OpenVoiceOS/nvidia-parakeet-ctc-1.1b-onnx at revision 3ca664a2f106622d599052b4e4ecee5fdfc7e2e5.

file	source	change
`model.int8.onnx`	`model.int8.onnx`	76 bytes of `metadata_props` appended
`tokens.txt`	`vocab.txt`	renamed only, contents unchanged

Added metadata

key	value
`vocab_size`	1025
`subsampling_factor`	8
`normalize_type`	per_feature

These are appended as three top-level metadata_props (ONNX ModelProto field 14) records. No weights, graph nodes, or existing metadata are touched — protobuf field ordering is not significant, so appending is equivalent to setting the keys.

Reproducing this artifact

bash

REV=3ca664a2f106622d599052b4e4ecee5fdfc7e2e5
BASE=https://huggingface.co/OpenVoiceOS/nvidia-parakeet-ctc-1.1b-onnx/resolve/$REV

curl -L -o model.int8.onnx "$BASE/model.int8.onnx"
curl -L -o tokens.txt      "$BASE/vocab.txt"

python3 -c "
payload = (
    '72120a0a766f6361625f73697a6512043130323572170a1273756273616d'
    '706c696e675f666163746f72120138721d0a0e6e6f726d616c697a655f74'
    '797065120b7065725f66656174757265'
)
open('model.int8.onnx', 'ab').write(bytes.fromhex(payload))
"

stage	size (bytes)	SHA-256
upstream `model.int8.onnx`	1110014069	`a16056c0a0d8df38c7b57cb019062df116e9e565203c6f25d6ea0c0c1122c84d`
this `model.int8.onnx`	1110014145	`62f73c17a5301c048c7273cf24ef1cd0c3621d3625c5415fbafe5633d7bf2f98`
`tokens.txt`	10374	`ed16e1a4e3a3aa379138c0b1888e5d49f993c9d512b2be4d46e90a87afd54921`

Graph interface

inputs : audio_signal, length
outputs: logprobs

License and attribution

CC-BY-4.0, inherited from nvidia/parakeet-ctc-1.1b. Original model © NVIDIA; INT8 ONNX export by the OpenVoiceOS project. This repository redistributes that work with the metadata fix described above.

runanywhere/sherpa-onnx-nemo-parakeet-ctc-1.1b-int8