| license: mit | |
| tags: | |
| - coreml | |
| - audio | |
| - speech | |
| - forced-alignment | |
| language: | |
| - en | |
| # Word Aligner: wav2vec 2.0 base for Core ML (int8) | |
| Times English words against the audio for Clarity Sound Studio, a free, open-source macOS app. The app aligns its transcript with this model's character outputs (CTC), so word cuts land on word boundaries. | |
| | | | | |
| | --- | --- | | |
| | File | `WordAligner-w2v2-base-en-int8.mlpackage` (ML Program, 8-bit weights, macOS 14+) | | |
| | Upstream | wav2vec 2.0 base fine-tuned on LibriSpeech 960 h, as redistributed by torchaudio (`torchaudio.pipelines.WAV2VEC2_ASR_BASE_960H`) | | |
| | Conversion | `convert_aligner.py --int8`: traced with torchaudio's model definition, converted with coremltools, weights quantised to 8 bits | | |
| The model is otherwise unchanged. | |
| ## Licence | |
| The weights are published by the fairseq authors under the MIT licence; see `LICENSE`. Copyright (c) Facebook, Inc. and its affiliates. | |
| > A. Baevski, H. Zhou, A. Mohamed and M. Auli, "wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations", NeurIPS 2020. <https://github.com/pytorch/fairseq/tree/ce6c9eeae163ac04b79539c78e74f292f29eaa18/examples/wav2vec> | |
| The conversion traced torchaudio's model definition, Copyright (c) 2017 Facebook Inc. (Soumith Chintala), BSD-2-Clause licence, <https://github.com/pytorch/audio>. No torchaudio code is in this repository. | |