spice.ipynb

Kernel: Python 3

Copyright 2020 The TensorFlow Hub Authors.

Licensed under the Apache License, Version 2.0 (the "License");

#@title Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================

在 TensorFlow.org 上查看

在 Google Colab 中运行

在 GitHub 上查看源代码

下载笔记本

看到 TF Hub models

使用 SPICE 进行音高检测

此 Colab 将为您展示如何使用从 TensorFlow Hub 下载的 SPICE 模型。

!sudo apt-get install -q -y timidity libsndfile1

# All the imports to deal with sound data
!pip install pydub librosa music21

import tensorflow as tf
import tensorflow_hub as hub

import numpy as np
import matplotlib.pyplot as plt
import librosa
from librosa import display as librosadisplay

import logging
import math
import statistics
import sys

from IPython.display import Audio, Javascript
from scipy.io import wavfile

from base64 import b64decode

import music21
from pydub import AudioSegment

logger = logging.getLogger()
logger.setLevel(logging.ERROR)

print("tensorflow: %s" % tf.__version__)
#print("librosa: %s" % librosa.__version__)

音频输入文件

下面是最难的部分：录制您的歌声！😃

我们提供四种获取音频文件的方法：

1. 直接在 Colab 中录制音频

2. 从您的计算机上传

3. 使用保存在 Google 云端硬盘上的文件

4. 从网上下载文件

在下面选择四种方法中的一种。

#@title [Run this] Definition of the JS code to record audio straight from the browser

RECORD = """
const sleep  = time => new Promise(resolve => setTimeout(resolve, time))
const b2text = blob => new Promise(resolve => {
  const reader = new FileReader()
  reader.onloadend = e => resolve(e.srcElement.result)
  reader.readAsDataURL(blob)
})
var record = time => new Promise(async resolve => {
  stream = await navigator.mediaDevices.getUserMedia({ audio: true })
  recorder = new MediaRecorder(stream)
  chunks = []
  recorder.ondataavailable = e => chunks.push(e.data)
  recorder.start()
  await sleep(time)
  recorder.onstop = async ()=>{
    blob = new Blob(chunks)
    text = await b2text(blob)
    resolve(text)
  }
  recorder.stop()
})
"""

def record(sec=5):
  try:
    from google.colab import output
  except ImportError:
    print('No possible to import output from google.colab')
    return ''
  else:
    print('Recording')
    display(Javascript(RECORD))
    s = output.eval_js('record(%d)' % (sec*1000))
    fname = 'recorded_audio.wav'
    print('Saving to', fname)
    b = b64decode(s.split(',')[1])
    with open(fname, 'wb') as f:
      f.write(b)
    return fname

#@title Select how to input your audio  { run: "auto" }
INPUT_SOURCE = 'https://storage.googleapis.com/download.tensorflow.org/data/c-scale-metronome.wav' #@param ["https://storage.googleapis.com/download.tensorflow.org/data/c-scale-metronome.wav", "RECORD", "UPLOAD", "./drive/My Drive/YOUR_MUSIC_FILE.wav"] {allow-input: true}

print('You selected', INPUT_SOURCE)

if INPUT_SOURCE == 'RECORD':
  uploaded_file_name = record(5)
elif INPUT_SOURCE == 'UPLOAD':
  try:
    from google.colab import files
  except ImportError:
    print("ImportError: files from google.colab seems to not be available")
  else:
    uploaded = files.upload()
    for fn in uploaded.keys():
      print('User uploaded file "{name}" with length {length} bytes'.format(
          name=fn, length=len(uploaded[fn])))
    uploaded_file_name = next(iter(uploaded))
    print('Uploaded file: ' + uploaded_file_name)
elif INPUT_SOURCE.startswith('./drive/'):
  try:
    from google.colab import drive
  except ImportError:
    print("ImportError: files from google.colab seems to not be available")
  else:
    drive.mount('/content/drive')
    # don't forget to change the name of the file you
    # will you here!
    gdrive_audio_file = 'YOUR_MUSIC_FILE.wav'
    uploaded_file_name = INPUT_SOURCE
elif INPUT_SOURCE.startswith('http'):
  !wget --no-check-certificate 'https://storage.googleapis.com/download.tensorflow.org/data/c-scale-metronome.wav' -O c-scale.wav
  uploaded_file_name = 'c-scale.wav'
else:
  print('Unrecognized input format!')
  print('Please select "RECORD", "UPLOAD", or specify a file hosted on Google Drive or a file from the web to download file to download')

准备音频数据

现在有了音频，我们将其转换为预期的格式，然后听听看！

SPICE 模型需要以采样率为 16kHz 且只有一个声道（单声道）的音频文件作为输入。

为了帮助您完成本部分，我们创建了一个函数（convert_audio_for_model），它可以将您手上的任何 WAV 文件转换为模型的预期格式：

# Function that converts the user-created audio to the format that the model 
# expects: bitrate 16kHz and only one channel (mono).

EXPECTED_SAMPLE_RATE = 16000

def convert_audio_for_model(user_file, output_file='converted_audio_file.wav'):
  audio = AudioSegment.from_file(user_file)
  audio = audio.set_frame_rate(EXPECTED_SAMPLE_RATE).set_channels(1)
  audio.export(output_file, format="wav")
  return output_file

# Converting to the expected format for the model
# in all the input 4 input method before, the uploaded file name is at
# the variable uploaded_file_name
converted_audio_file = convert_audio_for_model(uploaded_file_name)

# Loading audio samples from the wav file:
sample_rate, audio_samples = wavfile.read(converted_audio_file, 'rb')

# Show some basic information about the audio.
duration = len(audio_samples)/sample_rate
print(f'Sample rate: {sample_rate} Hz')
print(f'Total duration: {duration:.2f}s')
print(f'Size of the input: {len(audio_samples)}')

# Let's listen to the wav file.
Audio(audio_samples, rate=sample_rate)

首先，我们来看一下歌声的波形。

# We can visualize the audio as a waveform.
_ = plt.plot(audio_samples)

频谱图更具参考性，可以显示随时间变化的频率。

我们在这里使用对数频率标度，使歌声更加清晰可见。

MAX_ABS_INT16 = 32768.0

def plot_stft(x, sample_rate, show_black_and_white=False):
  x_stft = np.abs(librosa.stft(x, n_fft=2048))
  fig, ax = plt.subplots()
  fig.set_size_inches(20, 10)
  x_stft_db = librosa.amplitude_to_db(x_stft, ref=np.max)
  if(show_black_and_white):
    librosadisplay.specshow(data=x_stft_db, y_axis='log', 
                             sr=sample_rate, cmap='gray_r')
  else:
    librosadisplay.specshow(data=x_stft_db, y_axis='log', sr=sample_rate)

  plt.colorbar(format='%+2.0f dB')

plot_stft(audio_samples / MAX_ABS_INT16 , sample_rate=EXPECTED_SAMPLE_RATE)
plt.show()

我们需要在这里进行最后的转换。音频样本为 int16 格式。需要将它们归一化为 -1 和 1 之间的浮点数。

audio_samples = audio_samples / float(MAX_ABS_INT16)

执行模型

接下来的部分比较容易，我们来使用 TensorFlow Hub 加载模型，并将音频馈送到模型中。SPICE 将为我们提供两个输出：音高和不确定性。

TensorFlow Hub 是一个库，用于发布、发现和使用机器学习模型的可重用部分。有了它，您就可以轻松利用机器学习来解决您的挑战。

要加载模型，您只需 Hub 模块和指向模型的网址：

# Loading the SPICE model is easy:
model = hub.load("https://tfhub.dev/google/spice/2")

**注：**这里有一个有趣的细节，Hub 上的所有模型网址均可用来下载和阅读文档，因此，如果将浏览器指向该链接，不仅可以阅读有关如何使用模型的文档，还可以详细了解模型的训练方式。

加载模型并准备好数据后，我们需要 3 行来得到结果：

# We now feed the audio to the SPICE tf.hub model to obtain pitch and uncertainty outputs as tensors.
model_output = model.signatures["serving_default"](tf.constant(audio_samples, tf.float32))

pitch_outputs = model_output["pitch"]
uncertainty_outputs = model_output["uncertainty"]

# 'Uncertainty' basically means the inverse of confidence.
confidence_outputs = 1.0 - uncertainty_outputs

fig, ax = plt.subplots()
fig.set_size_inches(20, 10)
plt.plot(pitch_outputs, label='pitch')
plt.plot(confidence_outputs, label='confidence')
plt.legend(loc="lower right")
plt.show()

我们移除所有具有低置信度（置信度<0.9）的音高估计值，并绘制剩余估计值，使结果更易于理解。

confidence_outputs = list(confidence_outputs)
pitch_outputs = [ float(x) for x in pitch_outputs]

indices = range(len (pitch_outputs))
confident_pitch_outputs = [ (i,p)  
  for i, p, c in zip(indices, pitch_outputs, confidence_outputs) if  c >= 0.9  ]
confident_pitch_outputs_x, confident_pitch_outputs_y = zip(*confident_pitch_outputs)
 
fig, ax = plt.subplots()
fig.set_size_inches(20, 10)
ax.set_ylim([0, 1])
plt.scatter(confident_pitch_outputs_x, confident_pitch_outputs_y, )
plt.scatter(confident_pitch_outputs_x, confident_pitch_outputs_y, c="r")

plt.show()

SPICE 返回的音高值在 0 到 1 之间。我们将其转换为以 Hz 为单位的绝对音高值。

def output2hz(pitch_output):
  # Constants taken from https://tfhub.dev/google/spice/2
  PT_OFFSET = 25.58
  PT_SLOPE = 63.07
  FMIN = 10.0;
  BINS_PER_OCTAVE = 12.0;
  cqt_bin = pitch_output * PT_SLOPE + PT_OFFSET;
  return FMIN * 2.0 ** (1.0 * cqt_bin / BINS_PER_OCTAVE)
    
confident_pitch_values_hz = [ output2hz(p) for p in confident_pitch_outputs_y ]

现在，我们来看看预测的效果如何：我们将预测的音高叠加在原始频谱图上。为了使音高预测值更加清晰可见，我们将频谱图更改为黑白配色。

plot_stft(audio_samples / MAX_ABS_INT16 , 
          sample_rate=EXPECTED_SAMPLE_RATE, show_black_and_white=True)
# Note: conveniently, since the plot is in log scale, the pitch outputs 
# also get converted to the log scale automatically by matplotlib.
plt.scatter(confident_pitch_outputs_x, confident_pitch_values_hz, c="r")

plt.show()

转换为音符

现在有了音高值，我们来将它们转换为音符！这部分本身就是挑战。我们必须考虑以下两件事：

1. 剩余部分（没有歌声的部分）

2. 每个音符的大小（偏移）

1: 在输出中添加零来表示没有歌声的部分

pitch_outputs_and_rests = [
    output2hz(p) if c >= 0.9 else 0
    for i, p, c in zip(indices, pitch_outputs, confidence_outputs)
]

2: 添加音符偏移

当一个人自由歌唱时，旋律可能会相对音符代表的绝对音高值发生偏移。因此，为了将预测值转换为音符，需要校正这种可能的偏移。这就是下面的代码所计算的内容。

A4 = 440
C0 = A4 * pow(2, -4.75)
note_names = ["C", "C#", "D", "D#", "E", "F", "F#", "G", "G#", "A", "A#", "B"]

def hz2offset(freq):
  # This measures the quantization error for a single note.
  if freq == 0:  # Rests always have zero error.
    return None
  # Quantized note.
  h = round(12 * math.log2(freq / C0))
  return 12 * math.log2(freq / C0) - h


# The ideal offset is the mean quantization error for all the notes
# (excluding rests):
offsets = [hz2offset(p) for p in pitch_outputs_and_rests if p != 0]
print("offsets: ", offsets)

ideal_offset = statistics.mean(offsets)
print("ideal offset: ", ideal_offset)

现在，我们可以使用一些启发式方法尝试估计所演唱的最有可能的音符序列。上面计算出的理想偏移量只是其中一个因素，我们还需要知道速度（多少个预测值组成一个八分音符？），以及开始量化的时间偏移量。为简单起见，我们将尝试使用不同的速度和时间偏移量并测量量化误差，最后使用使该误差最小化的值。

def quantize_predictions(group, ideal_offset):
  # Group values are either 0, or a pitch in Hz.
  non_zero_values = [v for v in group if v != 0]
  zero_values_count = len(group) - len(non_zero_values)

  # Create a rest if 80% is silent, otherwise create a note.
  if zero_values_count > 0.8 * len(group):
    # Interpret as a rest. Count each dropped note as an error, weighted a bit
    # worse than a badly sung note (which would 'cost' 0.5).
    return 0.51 * len(non_zero_values), "Rest"
  else:
    # Interpret as note, estimating as mean of non-rest predictions.
    h = round(
        statistics.mean([
            12 * math.log2(freq / C0) - ideal_offset for freq in non_zero_values
        ]))
    octave = h // 12
    n = h % 12
    note = note_names[n] + str(octave)
    # Quantization error is the total difference from the quantized note.
    error = sum([
        abs(12 * math.log2(freq / C0) - ideal_offset - h)
        for freq in non_zero_values
    ])
    return error, note


def get_quantization_and_error(pitch_outputs_and_rests, predictions_per_eighth,
                               prediction_start_offset, ideal_offset):
  # Apply the start offset - we can just add the offset as rests.
  pitch_outputs_and_rests = [0] * prediction_start_offset + \
                            pitch_outputs_and_rests
  # Collect the predictions for each note (or rest).
  groups = [
      pitch_outputs_and_rests[i:i + predictions_per_eighth]
      for i in range(0, len(pitch_outputs_and_rests), predictions_per_eighth)
  ]

  quantization_error = 0

  notes_and_rests = []
  for group in groups:
    error, note_or_rest = quantize_predictions(group, ideal_offset)
    quantization_error += error
    notes_and_rests.append(note_or_rest)

  return quantization_error, notes_and_rests


best_error = float("inf")
best_notes_and_rests = None
best_predictions_per_note = None

for predictions_per_note in range(20, 65, 1):
  for prediction_start_offset in range(predictions_per_note):

    error, notes_and_rests = get_quantization_and_error(
        pitch_outputs_and_rests, predictions_per_note,
        prediction_start_offset, ideal_offset)

    if error < best_error:      
      best_error = error
      best_notes_and_rests = notes_and_rests
      best_predictions_per_note = predictions_per_note

# At this point, best_notes_and_rests contains the best quantization.
# Since we don't need to have rests at the beginning, let's remove these:
while best_notes_and_rests[0] == 'Rest':
  best_notes_and_rests = best_notes_and_rests[1:]
# Also remove silence at the end.
while best_notes_and_rests[-1] == 'Rest':
  best_notes_and_rests = best_notes_and_rests[:-1]

现在，我们将量化的音符编写成乐谱！

为此，我们将用到两个库：music21 和 Open Sheet Music Display

**注：**为简单起见，我们在此假设所有音符都具有相同的持续时间（二分音符）。

# Creating the sheet music score.
sc = music21.stream.Score()
# Adjust the speed to match the actual singing.
bpm = 60 * 60 / best_predictions_per_note
print ('bpm: ', bpm)
a = music21.tempo.MetronomeMark(number=bpm)
sc.insert(0,a)

for snote in best_notes_and_rests:   
    d = 'half'
    if snote == 'Rest':      
      sc.append(music21.note.Rest(type=d))
    else:
      sc.append(music21.note.Note(snote, type=d))

#@title [Run this] Helper function to use Open Sheet Music Display (JS code) to show a music score

from IPython.core.display import display, HTML, Javascript
import json, random

def showScore(score):
    xml = open(score.write('musicxml')).read()
    showMusicXML(xml)
    
def showMusicXML(xml):
    DIV_ID = "OSMD_div"
    display(HTML('<div id="'+DIV_ID+'">loading OpenSheetMusicDisplay</div>'))
    script = """
    var div_id = %%DIV_ID%%;
    function loadOSMD() { 
        return new Promise(function(resolve, reject){
            if (window.opensheetmusicdisplay) {
                return resolve(window.opensheetmusicdisplay)
            }
            // OSMD script has a 'define' call which conflicts with requirejs
            var _define = window.define // save the define object 
            window.define = undefined // now the loaded script will ignore requirejs
            var s = document.createElement( 'script' );
            s.setAttribute( 'src', "https://cdn.jsdelivr.net/npm/[email protected]/build/opensheetmusicdisplay.min.js" );
            //s.setAttribute( 'src', "/custom/opensheetmusicdisplay.js" );
            s.onload=function(){
                window.define = _define
                resolve(opensheetmusicdisplay);
            };
            document.body.appendChild( s ); // browser will try to load the new script tag
        }) 
    }
    loadOSMD().then((OSMD)=>{
        window.openSheetMusicDisplay = new OSMD.OpenSheetMusicDisplay(div_id, {
          drawingParameters: "compacttight"
        });
        openSheetMusicDisplay
            .load(%%data%%)
            .then(
              function() {
                openSheetMusicDisplay.render();
              }
            );
    })
    """.replace('%%DIV_ID%%',DIV_ID).replace('%%data%%',json.dumps(xml))
    display(Javascript(script))
    return

# rendering the music score
showScore(sc)
print(best_notes_and_rests)

我们将音符转换为 MIDI 文件听听看。

要创建此文件，我们可以使用之前创建的流。

# Saving the recognized musical notes as a MIDI file
converted_audio_file_as_midi = converted_audio_file[:-4] + '.mid'
fp = sc.write('midi', fp=converted_audio_file_as_midi)

wav_from_created_midi = converted_audio_file_as_midi.replace(' ', '_') + "_midioutput.wav"
print(wav_from_created_midi)

要在 Colab 上收听，我们需要将其转换回 WAV。一种简单的方法是使用 Timidity。

!timidity $converted_audio_file_as_midi -Ow -o $wav_from_created_midi

最后，听一下由音符创建的音频，该音频由模型推理，根据预测的音高通过 MIDI 创建！

Audio(wav_from_created_midi)