Niger-Volta-LTI/africomplings-audio-processing-labo

Audio Processing Tutorial # africomplings-audio-processing-labo Audio analysis tools for speech corpora — compute corpus statistics and benchmark ASR models. ## Scripts ### `africomplings_compute_corpus_stats.py` Compute audio statistics for a corpus of WAV files and generate summary plots. #### What it computes per file | Column | Source | Description | |--------|--------|-------------| | `Filename` | — | Basename of the WAV file | | `Sample_Rate` | `soundfile.read()` | Audio sample rate (Hz) | | `Duration_s` | `len(samples) / sr` | Duration in seconds | | `Loudness_LUFS` | `pyloudnorm.Meter.integrated_loudness()` | Integrated loudness (ITU-R BS.1770) | | `LRA_dB` | `pyloudnorm.Meter.loudness_range()` | Loudness Range (ITU-R BS.1770-4, G90-G10 percentile spread) | | `Pitch_Hz` | `librosa.yin(fmin=75, fmax=600)` | Median fundamental frequency of voiced frames | #### Plots generated 2×2 grid saved as PNG: - Duration histogram + KDE curve - Loudness (LUFS) histogram + KDE curve - LRA (dB) histogram + KDE curve - Loudness vs LRA scatter plot #### Usage ```bash # Scan audio/ directory, skip plots python africomplings_compute_corpus_stats.py -d audio/ --no-plot # Custom output file with plots python africomplings_compute_corpus_stats.py -d /path/to/audio -o my_stats.tsv ``` | Argument | Description | Default | |----------|-------------|---------| | `-d, --audio-dir` | Directory to scan for `.wav` files (recursive) | `./audio` | | `-o, --output` | Output TSV file path | `corpus_stats.tsv` | | `--no-plot` | Skip plotting, only write TSV | `False` | --- ### `bench_asr.py` Benchmark ASR models on a transcribed corpus — compare accuracy and speed. #### What it reports **Per-file table:** WER for each audio file per model. **Aggregate summary:** | Metric | Description | |--------|-------------| | **WER** (Word Error Rate) | 0 = perfect, lower is better | | **RTF** (Realtime Factor) | `< 1.0` means faster than real-time | | **Avg Latency** | Mean processing time per file | #### Usag …