turnover-engines

Model files and ffmpeg builds for an audio-analysis tool in beta.

Files in this repository

File What it is Licence
wespeaker-resnet34-LM-masked-fp16.onnx Speaker embedding model. An ONNX export of WeSpeaker ResNet34-LM, as wrapped in pyannote/wespeaker-voxceleb-resnet34-LM. Changes: pyannote's frame-weight (mask) input is kept; weights are stored as float16 and computed in float32; no retraining. Inputs: fbank (batch, 998, 80) and weights (batch, 589). Output: embedding (batch, 256). CC BY 4.0
calibration-ced-small-int8.json A table of monotone maps that rescales six summary scores of CED-small to another score scale. CC BY 4.0
manifest.json File list (version 1) with addresses, pinned revisions, sizes and SHA-256 checksums. CC BY 4.0
manifest-2026-10-06.json File list (version 2): the same model files plus the ffmpeg builds. Never edited once published. CC BY 4.0
ffmpeg/8.1.3-lgpl-1/ ffmpeg and ffprobe built from FFmpeg 8.1.3 with libsoxr 0.1.3, for Apple silicon (arm64/) and Intel (x86_64/). Complete source code and the build script are in source/. See NOTICE.txt and BUILD-INFO.txt in that folder. LGPL-2.1-or-later
MODEL-NOTICE.txt Attribution and licence notices for the model files. n/a

Files fetched from their original hosts

Listed in manifest.json, at pinned revisions; not copied here.

File Licence Host
CED-small, int8 ONNX Apache-2.0 k2-fsa/sherpa-onnx-ced-small-audio-tagging-2024-04-19, converted from mispeech/ced-small
AudioSet class labels CC BY 4.0 same repository
pyannote segmentation-3.0, ONNX MIT csukuangfj/sherpa-onnx-pyannote-segmentation-3-0, converted from pyannote/segmentation-3.0

ffmpeg

This repository includes ffmpeg and ffprobe, built from FFmpeg 8.1.3 and licensed under the GNU LGPL version 2.1 or later. They use libsoxr 0.1.3 (LGPL 2.1). The complete source code is in ffmpeg/8.1.3-lgpl-1/source/, with the script that built the binaries from it. FFmpeg is unmodified; only its build options were chosen (no network, no GPL or non-free parts).

Citations

  • WeSpeaker: Hongji Wang et al., Wespeaker: A research and production oriented speaker embedding learning toolkit, ICASSP 2023.
  • VoxCeleb: Arsha Nagrani et al., VoxCeleb: a large-scale speaker identification dataset, Interspeech 2017; Joon Son Chung et al., VoxCeleb2: Deep speaker recognition, Interspeech 2018.
  • pyannote: Alexis Plaquet and Hervé Bredin, Powerset multi-class cross entropy loss for neural speaker diarization, Interspeech 2023; Hervé Bredin, pyannote.audio 2.1 speaker diarization pipeline: principle, benchmark, and recipe, Interspeech 2023.
  • CED: Heinrich Dinkel et al., CED: Consistent ensemble distillation for audio tagging, ICASSP 2024.
  • AudioSet: Jort F. Gemmeke et al., Audio Set: An ontology and human-labeled dataset for audio events, ICASSP 2017.

These files are provided as-is, without warranty of any kind.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support