CoNeTTE is an audio captioning system, which generate a short textual description of the sound events in any audio file. The architecture and training are explained in the corresponding paper. The model has been developped by me (Étienne Labbé) during my PhD. A simple interface to test CoNeTTE is available on HuggingFace website.

Training

Requirements

• Intended for Ubuntu 20.04 only. Requires java < 1.13, ffmpeg, yt-dlp, and zip commands.
• Recommanded GPU: NVIDIA V100 with 32GB VRAM.
• WavCaps dataset might requires more than 2 TB of disk storage. Other datasets requires less than 50 GB.

Installation

By default, only the pip inference requirements are installed for conette. To install training requirements you need to use the following command:

python -m pip install conette[train]

If you already installed conette for inference, it is highly recommanded to create another environment before installing conette for training.

Download external models and data

These steps might take a while (few hours to download and prepare everything depending on your CPU, GPU and SSD/HDD).

First, download the ConvNeXt, NLTK and spacy models :

conette-prepare data=none default=true pack_to_hdf=false csum_in_hdf_name=false pann=false

Then download the 4 datasets used to train CoNeTTE :

common_args="data.download=true pack_to_hdf=true audio_t=resample_mean_convnext audio_t.pretrain_path=cnext_bl_75 post_hdf_name=bl pretag=cnext_bl_75"

conette-prepare data=audiocaps audio_t.src_sr=32000 ${common_args}
conette-prepare data=clotho audio_t.src_sr=44100 ${common_args}
conette-prepare data=macs audio_t.src_sr=48000 ${common_args}
conette-prepare data=wavcaps audio_t.src_sr=32000 ${common_args} datafilter.min_audio_size=0.1 datafilter.max_audio_size=30.0 datafilter.sr=32000

Train a model

CNext-trans (baseline) on CL only (~3 hours on 1 GPU V100-32G)

conette-train expt=[clotho_cnext_bl] pl=baseline

CoNeTTE on AC+CL+MA+WC, specialized for CL (~4 hours on 1 GPU V100-32G)

conette-train expt=[camw_cnext_bl_for_c,task_ds_src_camw] pl=conette

CoNeTTE on AC+CL+MA+WC, specialized for AC (~3 hours on 1 GPU V100-32G)

conette-train expt=[camw_cnext_bl_for_a,task_ds_src_camw] pl=conette

Note 1: any training using AC data cannot be exactly reproduced because a part of this data is deleted from the YouTube source, and I cannot share my own audio files. Note 2: paper results are averaged scores over 5 seeds (1234-1238). The default training only uses seed 1234.

Inference only (without training)

Installation

python -m pip install conette[test]

Usage with python

from conette import CoNeTTEConfig, CoNeTTEModel

config = CoNeTTEConfig.from_pretrained("Labbeti/conette")
model = CoNeTTEModel.from_pretrained("Labbeti/conette", config=config)

path = "/your/path/to/audio.wav"
outputs = model(path)
candidate = outputs["cands"][0]
print(candidate)

The model can also accept several audio files at the same time (list[str]), or a list of pre-loaded audio files (list[Tensor]). In this second case you also need to provide the sampling rate of this files:

import torchaudio

path_1 = "/your/path/to/audio_1.wav"
path_2 = "/your/path/to/audio_2.wav"

audio_1, sr_1 = torchaudio.load(path_1)
audio_2, sr_2 = torchaudio.load(path_2)

outputs = model([audio_1, audio_2], sr=[sr_1, sr_2])
candidates = outputs["cands"]
print(candidates)

The model can also produces different captions using a Task Embedding input which indicates the dataset caption style. The default task is "clotho".

outputs = model(path, task="clotho")
candidate = outputs["cands"][0]
print(candidate)

outputs = model(path, task="audiocaps")
candidate = outputs["cands"][0]
print(candidate)

Usage with command line

Simply use the command conette-predict with --audio PATH1 PATH2 ... option. You can also export results to a CSV file using --csv_export PATH.

conette-predict --audio "/your/path/to/audio.wav"

Performance

The model has been trained on AudioCaps (AC), Clotho (CL), MACS (MA) and WavCaps (WC). The performance on the test subsets are :

This model checkpoint has been trained with focus on the Clotho dataset, but it can also reach a good performance on AudioCaps with the "audiocaps" task.

Limitations

• The model expected audio sampled at 32 kHz. The model automatically resample up or down the input audio files. However, it might give worse results, especially when using audio with lower sampling rates.
• The model has been trained on audio lasting from 1 to 30 seconds. It can handle longer audio files, but it might require more memory and give worse results.

Citation

The preprint version of the paper describing CoNeTTE is available on arxiv: https://arxiv.org/pdf/2309.00454.pdf

@misc{labbe2023conette,
    title        = {CoNeTTE: An efficient Audio Captioning system leveraging multiple datasets with Task Embedding},
    author       = {Étienne Labbé and Thomas Pellegrini and Julien Pinquier},
    year         = 2023,
    journal      = {arXiv preprint arXiv:2309.00454},
    url          = {https://arxiv.org/pdf/2309.00454.pdf},
    eprint       = {2309.00454},
    archiveprefix = {arXiv},
    primaryclass = {cs.SD}
}

Additional information

• CoNeTTE stands for ConvNeXt-Transformer with Task Embedding.
• Model weights are available on HuggingFace: https://huggingface.co/Labbeti/conette
• The weights of the encoder part of the architecture is based on a ConvNeXt model for audio classification, available here: https://zenodo.org/records/10987498 under the filename "convnext_tiny_465mAP_BL_AC_75kit.pth".

Contact

Maintainer:

• Étienne Labbé "Labbeti": labbeti.pub@gmail.com