PitchBench

A benchmark for testing what audio / acoustic signals Audio Language Models (ALMs) do and don't understand. PitchBench probes pitch perception across 28 controlled experiments — single-pitch ID, onsets/offsets, chords, sequences, contour, audio effects, and polyphonic streams.

Each row is one (audio, question, answer) triple: a short WAV stimulus, the question (prompt*) asked of the model, and the ground-truth answer fields (experiment-specific column names).

Quick start

from datasets import load_dataset

# Load one experiment (configurations match experiment IDs)
ds = load_dataset("vaclis/PitchBench", "pitchbench_a1_pitch_id", split="test")
print(ds[0]["audio"], ds[0]["prompt_midi"], ds[0]["midi"])

# Iterate every experiment
import datasets
for cfg in datasets.get_dataset_config_names("vaclis/PitchBench"):
    ds = load_dataset("vaclis/PitchBench", cfg, split="test")
    print(cfg, len(ds))

Experiments (28 configs, 5,115 total stimuli)

Config	# rows	Question
pitchbench_a1_pitch_id	305	Identify the pitch of a single tone.
pitchbench_a2_pitch_with_reference	150	Identify a target pitch given a named reference tone.
pitchbench_a3_pitch_by_duration	700	Pitch ID across very short to very long tone durations.
pitchbench_a4_pitch_with_vibrato	140	Pitch ID with vibrato (rate × depth sweep).
pitchbench_a5_pitch_slightly_off	350	Pitch ID when the tone is detuned by a fraction of a semitone.
pitchbench_b1_pitch_in_silence	140	Pitch ID when the tone is hidden in a long silent stimulus.
pitchbench_b2_onset_offset_single	560	Predict the onset/offset times of a single tone in silence.
pitchbench_b3_onset_offset_specific	30	Predict onset/offset of a specific named target among distractors.
pitchbench_b4_pitch_at_time	75	Identify which pitch is sounding at a given timestamp.
pitchbench_b5_onset_offset_each	30	Predict onset/offset for every note in a sequence.
pitchbench_c1_dyad_interval	60	Identify the interval (in semitones) of a two-note dyad.
pitchbench_c2_chord_pitch_count	80	Count the number of simultaneous pitches in a chord.
pitchbench_c3_chord_pitch_id	100	List every pitch in a chord (dyad / triad / seventh).
pitchbench_c4_chord_quality	50	Classify chord quality (major, minor, dim, aug, 7th, sus, …).
pitchbench_d1_seq_pitch_count	70	Count the number of distinct pitches in a sequential passage.
pitchbench_d2_pitch_difference	110	Decide whether the second tone is higher or lower (cents-scale).
pitchbench_d3_interval_id_seq	120	Identify the interval between two sequentially played pitches.
pitchbench_d4_contour_discrete	100	Describe the up/down contour of a discrete-step melody.
pitchbench_d5_contour_continuous	20	Describe the contour of a continuous pitch glide.
pitchbench_d6_pitch_ranking	40	Rank N tones (small cents-scale differences) from low to high.
pitchbench_d7_seq_pitch_id	15	Transcribe every pitch in a melodic sequence.
pitchbench_e1_loudness	420	Pitch ID at varying loudness levels.
pitchbench_e2_audio_effects	490	Pitch ID under audio effects (reverb, EQ, clip, saturation, …).
pitchbench_e3_background_effects	100	Pitch ID embedded in real-world background noise (rain, crowd, …).
pitchbench_e4_harmonic_saturation	280	Pitch ID under increasing harmonic-saturation drive.
pitchbench_e5_time_stretch	300	Pitch ID with resample (pitch-shift) vs time-stretch (pitch preserved).
pitchbench_g1_melodic_line_id	255	Identify the pitch sequence of one part within a polyphonic mix.
pitchbench_g2_chorale_voice_id	25	Identify a target voice in a four-part Bach chorale rendering.

Schema

Every row has:

• file_name / audio — the WAV stimulus (16 kHz mono).
• prompt or one or more of prompt_midi, prompt_spn, prompt_abc, prompt_doremi, prompt_hz — the question(s) put to the model.
• Experiment-specific ground-truth fields (e.g. midi, n, interval_st, chord_quality_gt, pattern_gt, traj_name, midi_sequence, …).

Reproducibility

Stimuli are generated deterministically from the configuration in pitchbench.config (EVAL=True benchmark constants). The subset published here is the seeded stratified sample used in the paper — reproduced by apply_default_sampling(EXP_NAME, all_conds, None, seed=42) over the output of each experiment's build_conditions(...). Source code lives at github/PitchBench (insert the canonical link before public release).

License

Released under CC-BY-4.0.

Citation

@misc{pitchbench2026,
  title  = {PitchBench: A Benchmark for Pitch Understanding in Audio Language Models},
  author = {<authors>},
  year   = {2026},
}