kinomat-aligner-de

German speech model in ONNX, with the weights of its matrix multiplications quantized to int8, used by Kinomat for forced alignment: it finds when each word of a known German script is spoken in a voice recording, so that subtitles and cuts follow the voice.

This is a format conversion only. No weights were retrained or fine-tuned.

Deutsch: für Kinomat-Nutzer

Diese Datei herunterladen und in ~/Kinomat/Modelle/ ablegen: wav2vec2-base-german-cv9-int8.onnx. Den Dateinamen nicht ändern. Die Diagnose in Kinomat prüft danach die SHA-256-Prüfsumme unten. Kinomat lädt nichts selbst herunter.

File

File	wav2vec2-base-german-cv9-int8.onnx
Size	121,925,723 bytes
SHA-256	4a5bd72b104c3fc4e1a552c8e5bdb9f4135c4d01c81ad3b85af6775988aa65b0
Input	input_values: float32 [1, samples], 16 kHz mono, normalized to zero mean and unit variance
Output	logits: float32 [1, frames, 35], one frame per 20 ms
Vocabulary	vocab.json: lowercase German letters with ä, ö, ü, ß; the pipe character separates words, [PAD] is the CTC blank

Attribution

• Model: oliverguhr/wav2vec2-base-german-cv9 by Oliver Guhr, revision e3c2cb317c771e7fbbdfbf20be6017b8e65b232d, published under the MIT license (see LICENSE). The original model card declares the MIT license; it carries no separate license file or copyright line, so the copyright notice in LICENSE names the author without a year.
• Base model: facebook/wav2vec2-base by Meta AI, Apache License 2.0 (see LICENSE-wav2vec2-base).
• Training data of the original model: Mozilla Common Voice 9.0, German, CC0 1.0.

How this file was made

Exported from the original model at the revision above with exporter-le-modele.py, PyTorch 2.14.1, transformers 5.20.2, ONNX opset 17, then dynamic int8 quantization (uint8) of the MatMul weights only with onnxruntime 1.31.0. The convolutions of the feature encoder stay in float32: quantizing them too lowered frame agreement with the original model from 97 % to 82 % on a test recording.

python3.14 -m venv venv
venv/bin/pip install huggingface_hub==2.3.0 numpy==2.5.4 onnx==1.23.2 onnxruntime==1.31.0 onnxscript==0.7.2 torch==2.14.1 transformers==5.20.2
venv/bin/python -I exporter-le-modele.py out/

The export script is reproduced in this repository as exporter-le-modele.py. The exact bytes may differ between machines; the SHA-256 above is the one of the published file.

Limitations

Intended for aligning a known German script to a recording of that script. Not meant as a general speech recognizer. Provided as is, without warranty of any kind.