Search Similar Audio
Do you want to find similar songs from your music library? Do you want to search all the meeting records that your leader has given a comment? Do you want to find the timestamp from your videos when your baby is laughing? In all these cases, searching similar audios will be helpful. As an important format of storing information, searching audios is an essential part for managing multimedial data. In this tutorial, we will build an example using the VGGish model.
Build the Flow
Segment the Audio Clips
In this example, we use the AudioSet dataset. The dataset contains millions of annotated audios events extracted from YouTube videos. Each audio event is 10-second long and labeled to 632 audio event classes. One major challenges is that some audio events contains other events. This makes it difficult and noisy to express the whole clip with a single vector. For example, the audio clip below is labeled as Applause but contains a long part of music. To overcome this issue, we use the recursive structure of Jina Document and split the events into smaller chunks. Each chunk contains an audio clip of 4-second.
TipsThe AudioSet dataset doesn't contain the original audio event. You can use youtube-dl to download the audio data from the corresponding YouTube videos: youtube-dl --postprocessor-args '-ss 8.953 -to 18.953' -x --audio-format mp3 -o 'data/OXJ9Ln2sXJ8_30000.%(ext)s' https://www.youtube.com/watch\?v\=OXJ9Ln2sXJ8_30000
To segment the audio events into 4-second chunks, we create an executor, namely AudioSegmenter, which use librosa to load the audio files as waveform into blob. Afterwards, it splits the waveform array into smaller arrays based on the window_size. Each small array contains the audio data iin the wave form and is stored in the blob attribute of the chunks.
The stride argument is for setting the step size of the sliding window. Using stride=2 and window_size=4 to process a 10-second audio events, we will get 4 chunks, of which each chunk is 4-second long and has 2 seconds overlapped with the previous one.
AudioSegmenter requires the file path of the audio files to be defined in the uri attribute of each input Document.
import librosa as lr
class AudioSegmenter(Executor):
def __init__(self, window_size: float = 4, stride: float = 2, *args, **kwargs):
super().__init__(*args, **kwargs)
self.window_size = window_size # seconds
self.stride = stride
@requests(on=['/index', '/search'])
def segment(self, docs: DocumentArray, **kwargs):
for idx, doc in enumerate(docs):
try:
doc.blob, sample_rate = lr.load(doc.uri, sr=16000)
except RuntimeError as e:
print(f'failed to load {doc.uri}, {e}')
continue
doc.tags['sample_rate'] = sample_rate
chunk_size = int(self.window_size * sample_rate)
stride_size = int(self.stride * sample_rate)
num_chunks = max(1, int((doc.blob.shape[0] - chunk_size) / stride_size))
for chunk_id in range(num_chunks):
beg = chunk_id * stride_size
end = beg + chunk_size
if beg > doc.blob.shape[0]:
break
c = Document(
blob=doc.blob[beg:end],
offset=idx,
location=[beg, end],
tags=doc.tags,
uri=doc.uri
)
doc.chunksappend(c)
Notesample_rate is required for generating log mel spectrogram features and therefore we store this information at tags['sample_rate'].
TipsThe length of audios might not be exactly 10 seconds and therefore the number of extract chunks might vary from audio to audio.
Encode the Audio
To encode the sound clips into vectors, we choose VGGish model from Google Research. By default, the VGGish model needs the audios to be sampled at 16kHz and converted to examples of log mel spectrogram. The returning embeddings for each sound clip is a matrix of the size K x 128, where K is the number of examples in log mel spectrogram and roughly correspond to the length of audio in seconds. Therefore, each 4-second audio clip in the chunks is represented by four 128-dimensional vectors.
Since the sequence of the sounds matters, we further concatenate these four vectors and consider the resulted 512-dimensional vector as the final representation for each audio clip. After encoding indexing and querying audios into 512-dimensional vectors, we can find the similar audios to the querying ones by looking for nearest neighbors in the vector space.
VGGishAudioEncoder is available at Jina Hub. It accepts three types of inputs
- the waveform data stored in blob attribute together with the sampling rate information stored in tags['sample_rate']
- the log mel spectrogram features stored in blob attribute
- the file path information of a .mp3 or .wav file stored in the uri attribute
load_input_from argument is to configurate the input data type, which can be each waveform, log_mel, or uri. min_duration defines the number of vectors to concatenate.
...
- name: 'encoder'
uses: 'jinahub+docker://VGGishAudioEncoder/v0.4'
uses_with:
traversal_paths: ['c', ]
load_input_from: 'waveform'
min_duration: 4
volumes:
- './models:/workspace/models'
...
NoteWhen choosing waveform in VGGishAudioEncoder, we need to provide sample_rate at tags['sample_rate'] for generating log mel spectrogram features.
Storage
We choose the SimpleIndexer from Jina Hub for building a simple index storing both embedding vectors and meta information. During querying, we need to split the querying audios in the same way as indexing and generating chunks. Therefore, we need to set both traversal_rdarray and traversal_ldarray to ['c',] to ask the SimpleIndexer to use the embeddings of the chunks for the querying and the indexed Documents correspondingly.
...
- name: 'indexer'
uses: 'jinahub://SimpleIndexer/v0.7'
uses_with:
match_args:
limit: 5
traversal_rdarray: ['c',]
traversal_ldarray: ['c',]
...
Merge the Matches
Since we use audio chunks to retrieve the matches, we need to merge the retrieved matches into the matches for each query audio. We write MyRanker as below to collect the original 10-second audio event for each retrieved 4-second short clips. Since one audio event might be retrieved for multiple times base on different parts of its short clips, we use the score of the most matched short clip as the score of the audio event. Afterwards, the retrieved audio events are sorted by their scores.
class MyRanker(Executor):
@requests(on='/search')
def rank(self, docs: DocumentArray = None, **kwargs):
for doc in docs.traverse_flat(('r', )):
parents_scores = defaultdict(list)
parents_match = defaultdict(list)
for m in DocumentArray([doc]).traverse_flat(['cm']):
parents_scores[m.parent_id].append(m.scores['cosine'].value)
parents_match[m.parent_id].append(m)
new_matches = []
for match_parent_id, scores in parents_scores.items():
score_id = np.argmin(scores)
score = scores[score_id]
match = parents_match[match_parent_id][score_id]
new_match = Document(
uri=match.uri,
id=match_parent_id,
scores={'cosine': score})
new_matches.append(new_match)
# Sort the matches
doc.matches = new_matches
doc.matches.sort(key=lambda d: d.scores['cosine'].value)
Run the Flow
As we defined the flow in the YAML file, we use the load_config function to create the Flow and index the data.
from jina import DocumentArray, Flow
from jina.types.document.generators import from_files
docs = DocumentArray(from_files('toy-data/*.mp3'))
f = Flow.load_config('flow.yml')
with f:
f.post(on='/index', inputs=docs)
f.protocol = 'http'
f.cors = True
f.block()
Query from Python
With the Flow running as a http service, we can use the Jina swagger UI tool to query. Open the browser at localhost:45678/docs, send query via the Swagger UI,
{
"data": [
{
"uri": "toy-data/6pO06krKrf8_30000_airplane.mp3"
}
]
}
Show Results
| Query | Matches | Score |
|---|---|---|
| 0.002978 | ||
| 0.003295 | ||
| 0.000008 | ||
| 0.000011 |
Get the Source Code
The code is available at example-audio-search.
Your feedback helps us Improve!
We’d appreciate any feedback you’d have about your experience with the learning bootcamp. Please check it out and provide us with your valuable feedback.
