Cross-species F0 estimation, dataset and study of baseline algorithms

infos on current collaborations in https://docs.google.com/document/d/179dD1d6lmWhQ9e2E1AUoJyLZ1d_5c-aMNwAPkbmIcBU

metadata.py

Stores a dictionary of datasets and characteristics (SR, NFFT, path to access soundfiles, and downsampling factor for ultra/infra-sonic signals) Convenient to iterate over the whole dataset

from metadata import species
for specie in species:
    wavpath, FS, nfft, downsample, step = species[specie].values()
    # iterate over files (one per vocalisation)
    for fn in tqdm(glob(wavpath), desc=specie):
        sig, fs = sf.read(fn) # read soundfile
        annot = pd.read_csv(f'{fn[:-4]}.csv') # read annotations (one column Time in seconds, one column Freq in Herz)
        preds = pd.read_csv(f'{fn[:-4]}_preds.csv') # read the file gathering per algorithm f0 predictions

print_annot.py

For each vocalisation, prints a spectrogram and overlaid annotations as .png file stored in the annot_pngs folder.

run_all.py

Runs all baseline algorithms over the dataset.

• praat (praat-parselmouth implem)
• pyin (librosa implem)
• crepe (torchcrepe implem)
• crepe finetuned (torchcrepe implem)
• crepe finetuned over all species except the target
• crepe finetuned only on the target species
• crepe (original tensorflow implem https://arxiv.org/abs/1802.06182)
• basic pitch (https://arxiv.org/abs/2203.09893)
• pesto (https://arxiv.org/abs/2309.02265)
• pesto finetuned over the target species

This scripts stores predictions along with resampled annotations in {basename}_preds.csv files

print_annot.py

For each vocalisation, prints a spectrogram and overlaid annotations and predictions as .png file stored in the pred_pngs folder.

eval_all.py

Evaluates each algorithms over the dataset using {basename}_preds.csv files, with a threshold of 50 cents for accuracies. For each algorithms and species, this outputs ROC optimal thresholds, Recall, False alarm, Pitch accuracy, and Chroma accuracy. /!\ These metrics are mesured per vocalisation before being averaged. Scores are stored in scores/{specie}_scores.csv files

compute_salience_SHR.py

Evaluates metrics for each annotated temporal bin:

• the presence of a sub-harmonic following this paper
• the salience of the annotation as the ratio of the energy of the f0 (one tone around the annotation) and its surrounding (one octave around the annotation)
• the harmonicity of the annotation as the ratio between the energy of all harmonics and that of all harmonics except the fundamental These values are stored in the SHR, harmonicity and salience columns of the {basename}_preds.csv files

get_noisy_pngs.py

Thresholds salience and SHR values vocalisation wise and copies those considered "noisy" into the noisy_pngs folder to browse and check results. One can then check if the spectrogram exists in noisy_pngs to discard noisy labels.

train_crepe.py

Fine tunes the crepe model using the whole dataset.

• Loads 1024 sample windows and their corresponding f0 to be stored in a large train_set.pkl file (skip if data hasn't changed).
• Applies gradient descent using the BCE following the crepe paper (this task is treated as a binary classification for each spectral bin).
• The fine tuned model is stored in crepe_ft/model_all.pth
• Train on one target species given as argument (weights are stored in crepe_ft/model_only_{specie}.pth)
• Train on all species except the target given as argument (weights are stored in crepe_ft/model_omit_{specie}.pth)

Plotting

Scripts allow to generate plots to visualise results (they are saved as .pdf files in the figures folder)

• plot_freq_distrib.py generates a three panel subplot with violins showing distributions of f0 annotations in Hz, number of voiced bins per vocalisation, and modulation rate in (Hz/sec)
• plot_snr_distrib.py generates a three panel subplot with violins showing distributions of salience, SHR and harmonicity (see compute_salience_SHR.py)
• plot_scores_bars.py generates a 4 panel subplot showing performances for each species/algorithm combination. Arguments can be set to generate this plot over all vocalisations, sjipping vocalisations with a low salience / high SHR, or skipping time bins with a low salience / high SHR.