bass-transcribe

Bass Transcribe — played take → MIDI / Live clip / chart

Default for bass is the offline pyin pipeline. For vocals and expressive leads (sax, synth lead) switch to CREPE (§2b). For in-Live plugin options (NeuralNote, Jam Origin) read references/realtime-options.md.

Why pyin, not CREPE, for bass: in a side-by-side A/B in Live on a separated bass stem, the CREPE line was rejected by ear ("CREPE is not for bass") — it over-segments a slow-moving line into spurious adjacent notes. Published evidence agrees: pYIN beats CREPE on bass lines because CREPE's training set has almost no sub-bass examples (Araz, ISMIR 2021 LBD), and no public pitch benchmark covers the bass register. pyin is also ~30× faster. On vocals the same A/B went the other way — CREPE caught quiet, legato notes pyin dropped.

1. Get the audio

• User points at a file — wav/mp3/aiff of the take. Mono preferred; DI cleaner than mic.
• Record in Live — audio track, arm, record the take. MCP cannot export audio: the user does File → Export, or point at the raw recording under the Set's Samples/Recorded/ folder.
• Ask for (or detect) the take's BPM if the output will go into a Live clip.

2. Default pipeline — pyin → MIDI

import librosa, numpy as np, pretty_midi

y, sr = librosa.load(AUDIO, sr=22050, mono=True)
HOP = 256                                                    # 11.6 ms
f0, voiced, vprob = librosa.pyin(y, sr=sr, fmin=40.0, fmax=220.0,   # E1 … A3
                                 frame_length=4096, hop_length=HOP)
t = librosa.times_like(f0, sr=sr, hop_length=HOP)
rms = librosa.feature.rms(y=y, frame_length=4096, hop_length=HOP)[0]
rms = rms / (rms.max() + 1e-9)

VPROB, MIN_LEN, GAP = 0.55, 0.10, 0.05   # voicing gate, min note s, bridgeable dropout s
pitch = np.where(voiced & (vprob >= VPROB),
                 np.round(librosa.hz_to_midi(np.nan_to_num(f0, nan=40.0))), -1).astype(int)
notes, cur = [], None
for i, (ti, p) in enumerate(zip(t, pitch)):
    if cur and p == cur['p']:
        cur['end'], cur['i1'] = ti, i                        # same pitch: extend
        continue
    if cur and p < 0 and ti - cur['end'] <= GAP:
        continue                                             # short dropout: bridge it
    if cur and cur['end'] - cur['start'] >= MIN_LEN:
        notes.append(cur)                                    # pitch change / long gap: close
    cur = {'p': p, 'start': ti, 'end': ti, 'i0': i, 'i1': i} if p >= 0 else None
if cur and cur['end'] - cur['start'] >= MIN_LEN:
    notes.append(cur)

pm = pretty_midi.PrettyMIDI()
inst = pretty_midi.Instrument(program=33)                    # Electric Bass (finger)
for n in notes:
    vel = int(np.clip(50 + 70 * rms[n['i0']:n['i1'] + 1].mean(), 35, 120))
    inst.notes.append(pretty_midi.Note(velocity=vel, pitch=n['p'],
                                       start=n['start'], end=n['end']))
pm.instruments.append(inst)
pm.write(OUT_MID)