1

I've been working on a streaming audio fingerprinter in Python, and experimenting with a couple different libraries/services (Gracenote, AcoustID, ACRCloud), but can't get anything to work.

My current code uses python-sounddevice to record raw audio through my microphone (or soundcard for less noise) in a float32 format, and fill a ring buffer, which I then read every 5 seconds, convert to PCM 16 bit, and feed into the acoustid fingerprinting function.

My code:

import sounddevice as sd
import time as time2
duration = 15
b = Buffer(duration * RATE)
RATE = 16000
run_time = 60

class Buffer(object):
    def __init__(self, size, dtype=np.float32):
        self.size = size
        self.buf = np.zeros(self.size * 2, dtype=dtype)
        self.i = 0

    def extend(self, data):
        if len(data.shape) > 1:
            raise ValueError("data must be a flat array")

        l = data.size
        if l > self.size:
            raise ValueError("data cannot be larger than size")

        start = (self.i % self.size)
        end = start + l

        start_2 = start + self.size
        end_2 = end + self.size

        self.i += l

        if end < self.buf.size:
            self.buf[start:end] = data


        if end_2 < self.buf.size:
            self.buf[start_2:end_2] = data


    def read(self):
        start = (self.i % self.size)
        end = start + self.size

        return self.buf[start:end]

def float_to_16_bit_pcm(raw_floats):
    floats = array.array('f', raw_floats)
    samples = [sample * 32767 for sample in floats]
    raw_ints = struct.pack("<%dh" % len(samples), *samples)
    return raw_ints

def callback(indata, frames, time, status): #outdata is 5th - when no inputstream

    global run_time
    global i
    if status:
        print(status)

    b.extend(indata.squeeze())
    elapsed_time = time2.time()- start


    if elapsed_time > duration and i % 50 == 0:

        aud = b.read()
        pcm16 = float_to_16_bit_pcm(aud)
        fp = acoustid.fingerprint(16000, 1, pcm16)
        response = acoustid.lookup(API_KEY, fp, 15)
        print response
    i += 1
with sd.InputStream(samplerate=16000, dtype= np.float32, channels=1, callback=callback):

    sd.sleep(int(run_time*1000))

The response comes back with the error: {u'status': u'error', u'error': {u'message': u'invalid fingerprint', u'code': 3}}

I know that the creator of Chromaprint did mention that it is now possible to feed in a continuous audio stream to Chromaprint 1.4 (in C++) and get fingerprints here: https://oxygene.sk/2016/12/chromaprint-1-4-released/

Does anyone have any experience with this or advice to offer?

Thanks

Aaron Oppenheimer
  • 31
  • 2
  • The problem is that you are trying to use these kind of fingerprints with AcoustID. That is not supported. It was only added to Chromaprint for uses outside of AcoustID (e.g. matching directly in the client, or custom server). – Lukáš Lalinský Nov 11 '17 at 11:46
  • Ok, so there is no way to use raw, recorded audio as an input to the AcoustID algorithm? – Aaron Oppenheimer Nov 13 '17 at 18:51
  • You can pass it to the fingerprinting algorithm, but you can't use such fingerprint in the AcoustID service. – Lukáš Lalinský Nov 14 '17 at 17:15

0 Answers0