khoben/whistle-small-int8

Whistle Small (ONNX INT8)

A standalone ONNX export of the Whistle Small phoneme CTC acoustic model with dynamic INT8 weight quantization. It takes acoustic features and produces frame-level log-probabilities over phoneme CTC classes. These scores can support phonetic decoding and transcript-constrained alignment research; word transcription and word boundaries require additional processing.

The original model is by Saierdaer Yusuyin, Te Ma, Hao Huang, Wenbo Zhao and Zhijian Ou (THU-SPMI, Tsinghua University). This repository distributes a converted derivative and does not claim authorship or training of the original model. It is an unofficial redistribution.

Files

File	Bytes	SHA-256
whistle-small.int8.onnx	100234902	04452b16c79698b48ae846d520b3ef6d54df26f022ac6c186599440908542d69

SHA256SUMS records the model and licensing-file digests. Pin a repository commit in download URLs (resolve/<commit>/<file>) and verify size and SHA-256 after download.

Acoustic and phonetic capabilities

The upstream model card at the source revision describes a multilingual Conformer encoder trained with CTC and weak phonetic supervision from IPA transcriptions on Common Voice CV-Lang10. Its ten training languages are English, Spanish, French, Italian, Kyrgyz, Dutch, Russian, Swedish, Turkish and Tatar. Those upstream results do not establish the performance of this INT8 export.

The artifact emits 75 CTC class scores per output frame. Class scores alone do not specify their phoneme-to-index mapping. A compatible feature extractor and upstream token mapping, including the CTC blank index, are required. Word decoding additionally requires suitable decoding resources; forced alignment requires a transcript, a compatible phonetic front end and an alignment algorithm. Those resources are not included here.

Model interface

The following interface was inspected directly in the distributed ONNX graph:

• Inputs: feats float32 [1, T, 80]; lens int64 [1].
• Outputs: logprobs float32 [batch, To, 75], where the graph declares a symbolic batch dimension and the input batch size is 1; olens float32 [1].
• Standard ONNX opset 17. Symbolic time dimensions T and To.
• The graph contains LogSoftmax, DynamicQuantizeLinear, MatMulInteger and ConvInteger operations, and no external initializer files.
• The existing conversion record describes the input as Kaldi-style log-mel filterbank features. The exact preprocessing configuration is not included in this repository.

Conversion and provenance

The existing distribution record identifies these sources:

• Checkpoint: best-3.pt from THU-SPMI/whistle-small, revision 83add32935aeefac3a5fed910283e975fe946116 (Apache-2.0).
• Architecture: ConformerNet from the upstream THU-SPMI/CAT toolkit, commit 9b8163ec615db2598a5c14fab6722765e03f18b5 (Apache-2.0).

The recorded conversion loaded that checkpoint into ConformerNet, added log_softmax, and exported with torch.onnx.export at opset 17 with dynamic time axes. The recorded quantization used onnxruntime.quantization.quantize_dynamic, weight_type=QuantType.QInt8 and per_channel=True. The record states that the model was not retrained or fine-tuned and that its training data, architecture and vocabulary were unchanged apart from export and quantization.

Provenance gaps: This repository does not include the conversion script, a reproducible environment with dependency versions, the source checkpoint checksum, the intermediate FP32 ONNX graph or its checksum, or export/quantization logs. The graph's producer metadata is onnx.quantize version 0.1.0 with onnx.infer=onnxruntime.quant; this does not establish the exact PyTorch or ONNX Runtime versions. The historical conversion record cannot be fully reproduced or independently verified from the files here. The SHA-256 identifies the distributed artifact, not a proof of every recorded conversion step.

Limitations

• This export has no independent per-language recognition, phoneme-error or alignment benchmark in this repository. Upstream benchmark numbers should not be attributed to this artifact.
• INT8 quantization can change numerical outputs and downstream results relative to the FP32 export.
• The model accepts precomputed features; it does not directly consume a waveform or output words or word boundaries.
• Feature extraction, class ordering, blank handling and frame-to-time mapping must be validated for the intended decoder or alignment procedure. Output frames are not ground-truth word timings.

Attribution and license

Distributed under the Apache License, Version 2.0. See LICENSE and NOTICE for attribution, source revisions and modification notices. Redistribution must comply with Apache-2.0, including providing the license, retaining applicable notices and identifying modifications as required. The model binary and license are unchanged in this documentation revision. The redistribution wording in NOTICE has been revised, and the original upstream author attribution and modification disclosure are retained.

This is not an official THU-SPMI release, and THU-SPMI does not endorse it.