| --- |
| library_name: coreml |
| license: mit |
| tags: |
| - coreml |
| - speaker-verification |
| - speaker-embedding |
| - diarization |
| - redimnet |
| - audio |
| pipeline_tag: audio-classification |
| --- |
| |
| # ReDimNet2-B6 Core ML Speaker Embeddings |
|
|
| This directory contains a Core ML conversion of the ReDimNet2-B6 speaker embedding model from [`PalabraAI/redimnet2`](https://github.com/PalabraAI/redimnet2). |
|
|
| The model is used by software to assign deterministic speaker labels inside each audio file and prefix transcriptions with markers such as: |
|
|
| ```text |
| {SPEAKER_1} Добрий день. |
| {SPEAKER_2} Вітаю. |
| ``` |
|
|
| ## Files |
|
|
| ```text |
| ReDimNet2-B6.mlpackage/ |
| ``` |
|
|
| ## Model Details |
|
|
| - Source model: ReDimNet2-B6 |
| - Upstream repository: `PalabraAI/redimnet2` |
| - Checkpoint: `b6-vb2+vox2_v0-lm.pt` |
| - Task: speaker embedding extraction |
| - Input: mono 16 kHz waveform |
| - Output: L2-normalized speaker embedding |
| - Core ML input name: `audio` |
| - Core ML output name: `embedding` |
|
|
| The converted package expects a fixed waveform input of `160320` samples, about `10.02s` at 16 kHz. The software pads shorter chunks and center-crops longer chunks before inference. |
|
|
| ## Convert |
|
|
| From the repository root: |
|
|
| ```bash |
| uv run --with torch --with torchaudio --with scipy --with coremltools \ |
| scripts/convert_redimnet2_coreml.py \ |
| --output Models/speaker/ReDimNet2-B6.mlpackage |
| ``` |
|
|
| ## Notes |
|
|
| The model produces embeddings, not speaker IDs. The software performs per-file online cosine clustering over chunk embeddings. Speaker labels are deterministic within a source audio file, but `SPEAKER_1` in one file is not the same person as `SPEAKER_1` in another file. |
|
|
