Skip to content

A lightweight yet powerful audio-to-MIDI converter with pitch bend detection

add ffmpeg to github actions, pyproject, try to use in slakh

Benjie Genchel9914d6a1bdd4
diff --git a/.github/workflows/tox.yml b/.github/workflows/tox.yml
index e94459f..b06508a 100644
--- a/.github/workflows/tox.yml
+++ b/.github/workflows/tox.yml
@@ -30,13 +30,13 @@ jobs:
python-version: ${{ matrix.py }}
- uses: actions/checkout@v3
- name: Install soundlibs Ubuntu
- run: sudo apt-get update && sudo apt-get install --no-install-recommends -y --fix-missing pkg-config libsndfile1 sox
+ run: sudo apt-get update && sudo apt-get install --no-install-recommends -y --fix-missing pkg-config libsndfile1 sox ffmpeg
if: matrix.os == 'ubuntu-latest'
- name: Install soundlibs MacOs
- run: brew install libsndfile llvm libomp sox
+ run: brew install libsndfile llvm libomp sox ffmpeg
if: matrix.os == 'macos-latest-xlarge'
- name: Install soundlibs Windows
- run: choco install libsndfile sox.portable flac
+ run: choco install libsndfile sox.portable flac ffmpeg
if: matrix.os == 'windows-latest'
- name: Upgrade pip
run: python -m pip install -U pip
diff --git a/basic_pitch/data/datasets/slakh.py b/basic_pitch/data/datasets/slakh.py
index 6801cdb..260a77b 100644
--- a/basic_pitch/data/datasets/slakh.py
+++ b/basic_pitch/data/datasets/slakh.py
@@ -44,7 +44,7 @@ class SlakhFilterInvalidTracks(beam.DoFn):
import tempfile
import apache_beam as beam
- import sox
+ import ffmpeg
from basic_pitch.constants import (
AUDIO_N_CHANNELS,
@@ -77,11 +77,10 @@ class SlakhFilterInvalidTracks(beam.DoFn):
return None
local_wav_path = "{}_tmp.wav".format(track_local.audio_path)
- tfm = sox.Transformer()
- tfm.rate(AUDIO_SAMPLE_RATE)
- tfm.channels(AUDIO_N_CHANNELS)
try:
- tfm.build(track_local.audio_path, local_wav_path)
+ ffmpeg.input(track_local.audio_path).output(
+ local_wav_path, ar=AUDIO_SAMPLE_RATE, ac=AUDIO_N_CHANNELS
+ ).run()
except Exception as e:
logging.info(f"Could not process {local_wav_path}. Exception: {e}")
return None
@@ -113,7 +112,7 @@ class SlakhToTfExample(beam.DoFn):
import tempfile
import numpy as np
- import sox
+ import ffmpeg
from basic_pitch.constants import (
AUDIO_N_CHANNELS,
@@ -145,12 +144,11 @@ class SlakhToTfExample(beam.DoFn):
d.write(s.read())
local_wav_path = "{}_tmp.wav".format(track_local.audio_path)
- tfm = sox.Transformer()
- tfm.rate(AUDIO_SAMPLE_RATE)
- tfm.channels(AUDIO_N_CHANNELS)
- tfm.build(track_local.audio_path, local_wav_path)
+ ffmpeg.input(track_local.audio_path).output(
+ local_wav_path, ar=AUDIO_SAMPLE_RATE, ac=AUDIO_N_CHANNELS
+ ).run()
- duration = sox.file_info.duration(local_wav_path)
+ duration = float(ffmpeg.probe(local_wav_path)["format"]["duration"])
time_scale = np.arange(0, duration + ANNOTATION_HOP, ANNOTATION_HOP)
n_time_frames = len(time_scale)
diff --git a/pyproject.toml b/pyproject.toml
index 4e23b0f..8843074 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -57,7 +57,8 @@ data = [
"apache_beam",
"mirdata",
"smart_open",
- "sox"
+ "sox",
+ "ffmpeg-python"
]
test = [
"basic_pitch[data]",