| --- |
| license: apache-2.0 |
| language: en |
| tags: |
| - onnx |
| - wav2vec2 |
| - forced-alignment |
| base_model: facebook/wav2vec2-base-960h |
| --- |
| |
| # wav2vec2-base-960h, ONNX, 8-bit (word alignment) |
|
|
| Part of [Transcut Studio](https://transcutstudio.com), a video editor for Mac that edits video by editing text. |
|
|
| A converted copy of [facebook/wav2vec2-base-960h](https://huggingface.co/facebook/wav2vec2-base-960h) |
| (Apache-2.0), used by [Transcut Studio](https://transcutstudio.com) to place each word of a transcript precisely in the audio. |
|
|
| - Exported to ONNX (opset 17). Input `input_values` `[1, samples]`: 16 kHz mono, zero mean and |
| unit variance. Output `logits` `[1, frames, 32]`, one frame per 20 ms, the model's 32 symbols. |
| - The transformer's matrix multiplications are quantized to 8 bits (per channel); the |
| convolutional front-end is kept at 32 bits. Word times match the full model's. |
| - SHA-256: `79cf3a25b88b84660deaf48f178077fda2de3b5d793d0b14986a74e95ac6b4c9` |
|
|
| Licence: Apache-2.0 (see `LICENSE` and `NOTICE`). Original model © Facebook, Inc. and its affiliates. |
|
|
