Chroma Features (synctoolbox.feature.chroma)

synctoolbox.feature.chroma.pitch_to_CENS(f_pitch: ndarray, input_feature_rate: float, win_len_smooth: int = 0, downsamp_smooth: int = 1, quant_steps: ndarray = np.array([40, 20, 10, 5]) / 100, quant_weights: ndarray = np.array([1, 1, 1, 1]) / 4, norm_thresh: float = 0.001, midi_min: int = 21, midi_max: int = 108) → Tuple[ndarray, float][source]

Generate CENS features from pitch features (CENS: Chroma Energy Normalized Statistics).

The following is computed:

  • Energy for each chroma band

  • l1-normalization of the chroma vectors

  • Local statistics:

    • Component-wise quantization of the normalized chroma vectors

    • Smoothing and downsampling of the feature sequence

    • l2-normalization of the resulting vectors

Individual steps of this procedure can be computed with the remaining functions in this module.

Parameters

  • f_pitch (np.ndarray [shape=(128, N)]) – MIDI pitch-based feature representation, obtained e.g. through audio_to_pitch_features.

  • input_feature_rate (float) – Feature rate of the input pitch features f_pitch

  • win_len_smooth (int) – Smoothing window length, default: no smoothing

  • downsamp_smooth (int) – Downsampling factor, default: no downsampling

  • quant_steps (np.ndarray) – After l1-normalization, all entries are quantized into bins defined by these boundaries. The default values correspond to the standard definition of CENS features.

  • quant_weights (np.ndarray) – The individual quantization bins can be given weights. Default is equal weight for all bins.

  • norm_thresh (float) – For l1-normalization, chroma entries below this threshold are considered as noise and set to 0. For l2-normalization, chroma vectors with norm below this threshold are replaced with uniform vectors.

  • midi_min (int) – Minimum MIDI pitch index to consider (default: 21)

  • midi_max (int) – Maximum MIDI pitch index to consider (default: 108)

Returns

  • f_CENS (np.ndarray) – CENS (Chroma Energy Normalized Statistics) features

  • CENS_feature_rate (float) – Feature rate of the CENS features

synctoolbox.feature.chroma.pitch_to_chroma(f_pitch: ndarray, midi_min: int = 21, midi_max: int = 108) → ndarray[source]

Aggregate pitch-based features into chroma bands.

Parameters

  • f_pitch (np.ndarray [shape=(128, N)]) – MIDI pitch-based feature representation, obtained e.g. through audio_to_pitch_features.

  • midi_min (int) – Minimum MIDI pitch index to consider (default: 21)

  • midi_max (int) – Maximum MIDI pitch index to consider (default: 108)

Returns

f_chroma (np.ndarray [shape=(12, N)]) – Rows of ‘f_pitch’ between midi_min and midi_max, aggregated into chroma bands.

synctoolbox.feature.chroma.quantize_chroma(f_chroma, quant_steps: ndarray = np.array([40, 20, 10, 5]) / 100, quant_weights: ndarray = np.array([1, 1, 1, 1]) / 4, norm_thresh: float = 0.001) → ndarray[source]

Computes thresholded l1-normalization of the chroma vectors and then applies component-wise quantization of the normalized chroma vectors.

Parameters

  • f_chroma (np.ndarray [shape=(12, N)]) – Chroma representation

  • quant_steps (np.ndarray) – After l1-normalization, all entries are quantized into bins defined by these boundaries. The default values correspond to the standard definition of CENS features.

  • quant_weights (np.ndarray) – The individual quantization bins can be given weights. Default is equal weight for all bins.

  • norm_thresh (float) – For l1-normalization, chroma entries below this threshold are considered as noise and set to 0.

Returns

f_chroma_quantized (np.ndarray [shape=(12, N)]) – Quantized chroma representation

synctoolbox.feature.chroma.quantized_chroma_to_CENS(f_chroma_quantized: ndarray, win_len_smooth: int, downsamp_smooth: int, input_feature_rate: float, norm_thresh: float = 0.001)[source]

Smooths, downsamples, and normalizes a chroma sequence obtained e.g. through quantize_chroma.

Parameters

  • f_chroma_quantized (np.ndarray [shape=(12, N)]) – Quantized chroma representation

  • win_len_smooth (int) – Smoothing window length. Setting this to 0 applies no smoothing.

  • downsamp_smooth (int) – Downsampling factor. Setting this to 1 applies no downsampling.

  • input_feature_rate (float) – Feature rate of f_chroma_quantized

  • norm_thresh (float) – For the final l2-normalization, chroma vectors with norm below this threshold are replaced with uniform vectors.

Returns

  • f_CENS (np.ndarray) – CENS (Chroma Energy Normalized Statistics) features

  • CENS_feature_rate (float) – Feature rate of the CENS features