branchkit/nemo-ctc-en-conformer-small-onnx

NeMo Conformer-CTC small (English), ONNX for sherpa-onnx

An ONNX conversion of NVIDIA's nvidia/stt_en_conformer_ctc_small (~13M parameters, 4x subsampling, 1024-piece SentencePiece tokenizer), for offline CTC decoding with sherpa-onnx. It is the command-recognition model for BranchKit's voice plugin.

Attribution and license

"stt_en_conformer_ctc_small" by NVIDIA, https://huggingface.co/nvidia/stt_en_conformer_ctc_small (revision e5b9941cc1b0b8a08c29b31a111c674f3040a80f), is licensed under CC BY 4.0. These files are a modified version of it, distributed under the same license. They are provided as is, without warranties of any kind.

Changes made:

• model.onnx: the checkpoint (stt_en_conformer_ctc_small.nemo, sha256 f9af79844c7160b3668f373cd62dbf28b69df0e2e441c3b5e0f20159dd8c7088) restored with NeMo 2.1.0 as EncDecCTCModelBPE and exported with its own export(). sherpa-onnx metadata was added: vocab_size, normalize_type=per_feature, subsampling_factor=4, model_type=EncDecCTCModelBPE.
• model.int8.onnx: model.onnx with dynamic int8 quantization of MatMul weights (onnxruntime quantize_dynamic, QUInt8).
• tokens.txt: the model's decoder vocabulary, one piece id per line, with <blk> appended last.
• bpe.model: the checkpoint's SentencePiece tokenizer, unmodified.
• bpe.vocab: a text dump of bpe.model, one piece score per line.

Weights were not retrained or fine-tuned.

Files

file	size	sha256
model.onnx	60.1 MB	1088fc73ce8bd7dc39ccd72e36693bb574c47959fb5b60c930596f9daa255fbb
model.int8.onnx	27.4 MB	06828c7bc831e3d1be3a9c4aedcfd82d947e70e536dd2b01eb4e06712b462b86
tokens.txt	12 KB	1adce7866285e6a4aac6208b120d5b8e4302b7cae5f2ae541c343097773fd988
bpe.model	260 KB	d9b04033136c5d0413047fe94d2f0ab6cb088d292014bb076ee1700bdac545a9
bpe.vocab	27 KB	13d12e8b0fce5ad809774f53ddefb949b0b367b7217ec2e9875d32570030a095

Use with sherpa-onnx

import sherpa_onnx
rec = sherpa_onnx.OfflineRecognizer.from_nemo_ctc(model="model.int8.onnx", tokens="tokens.txt")

bpe.model and bpe.vocab are for grammar-constrained decoding, which spells arbitrary words into the model's tokens; plain decoding needs only the model and tokens.txt.