PyAudio Input overflowed

Question

I'm trying to make real-time plotting sound in python. I need to get chunks from my microphone.

Using PyAudio, try to use

import pyaudio
import wave
import sys

chunk = 1024
FORMAT = pyaudio.paInt16
CHANNELS = 1
RATE = 44100
RECORD_SECONDS = 5
WAVE_OUTPUT_FILENAME = "output.wav"

p = pyaudio.PyAudio()

stream = p.open(format = FORMAT,
                channels = CHANNELS,
                rate = RATE,
                input = True,
                frames_per_buffer = chunk)

print "* recording"
all = []
for i in range(0, RATE / chunk * RECORD_SECONDS):
    data = stream.read(chunk)
    all.append(data)
print "* done recording"

stream.close()
p.terminate()

After, I getting the followin error:

* recording
Traceback (most recent call last):
  File "gg.py", line 23, in <module>
    data = stream.read(chunk)
  File "/usr/lib64/python2.7/site-packages/pyaudio.py", line 564, in read
    return pa.read_stream(self._stream, num_frames)
IOError: [Errno Input overflowed] -9981

I can't understand this buffer. I want, to use blocking IO mode, so if chunks not available, i want to wait for those chunks. But when I creating try except segment or sleep(0.1), i hear clicks, so this is not what i want.

Please suggest the best solution for my ploblem?

Lyall

Perhaps your chunk size is too small. Maybe it is getting more data in the buffer than you are pulling out because the chunk size is small enough the Python code is not keeping up. – Diner

Hi. Just wondering if there are any updates on this issue? I am getting the [Errno Input overflowed] -9981 error intermittently. I have checked p.is_format_supported is true for the format I am using. – Skindive

Answer 1

It seems like a lot of people are encountering this issue. I dug a bit into it and I think it means that between the previous call to stream.read() and this current call, data from the stream was lost (i.e. the buffer filled up faster than you cleared it).

From the doc for Pa_ReadStream() (the PortAudio function that stream.read() eventually ends up calling):

@return On success PaNoError will be returned, or PaInputOverflowed if
input data was discarded by PortAudio after the previous call and
before this call.

(PaInputOverflowed then causes an IOError in the pyaudio wrapper).

If it's OK for you to not capture every single frame, then you may ignore this error. If it's absolutely critical for you to have every frame, then you'll need to find a way to increase the priority of your application. I'm not familiar enough with Python to know a pythonic way to do this, but it's worth trying a simple nice command, or changing the scheduling policy to SCHED_DEADLINE.

Edit:

One issue right now is that when IOError is thrown, you lose all the frames collected in that call. To instead ignore the overflow and just return what we have, you can apply the patch below, which will cause stream.read() to ignore output underrun and input overflow errors from PortAudio (but still throw something if a different error occurred). A better way would be to make this behaviour (throw/no throw) customizable depending on your needs.

diff --git a/src/_portaudiomodule.c b/src/_portaudiomodule.c
index a8f053d..0878e74 100644
--- a/src/_portaudiomodule.c
+++ b/src/_portaudiomodule.c
@@ -2484,15 +2484,15 @@ pa_read_stream(PyObject *self, PyObject *args)
     } else {
       /* clean up */
       _cleanup_Stream_object(streamObject);
+
+      /* free the string buffer */
+      Py_XDECREF(rv);
+
+      PyErr_SetObject(PyExc_IOError,
+                       Py_BuildValue("(s,i)",
+                                     Pa_GetErrorText(err), err));
+      return NULL;
     }
-
-    /* free the string buffer */
-    Py_XDECREF(rv);
-
-    PyErr_SetObject(PyExc_IOError,
-                   Py_BuildValue("(s,i)",
-                                 Pa_GetErrorText(err), err));
-    return NULL;
   }

   return rv;

Spathose

Answer 2

pyaudio.Stream.read() has a keyword parameter exception_on_overflow, set this to False.

For your sample code that would look like:

import pyaudio
import wave
import sys

chunk = 1024
FORMAT = pyaudio.paInt16
CHANNELS = 1
RATE = 44100
RECORD_SECONDS = 5
WAVE_OUTPUT_FILENAME = "output.wav"

p = pyaudio.PyAudio()

stream = p.open(format = FORMAT,
                channels = CHANNELS,
                rate = RATE,
                input = True,
                frames_per_buffer = chunk)

print "* recording"
all = []
for i in range(0, RATE / chunk * RECORD_SECONDS):
    data = stream.read(chunk, exception_on_overflow = False)
    all.append(data)
print "* done recording"

stream.close()
p.terminate()

See the PyAudio documentation for more details.

Trunkfish

I get: TypeError: read() got an unexpected keyword argument 'exception_on_overflow' – Japonica

Answer 3

I got the same error when I ran your code. I looked at the default sample rate of my default audio device, my macbook's internal microphone, it was 48000Hz not 44100Hz.

p.get_device_info_by_index(0)['defaultSampleRate']
Out[12]: 48000.0

When I changed RATE to this value, it worked.

Agatha

I got the same error, and your solution (upping to 48000) worked. But I had run the code: if p.is_format_supported(44100.0, # Sample rate input_device=devinfo["index"], input_channels=devinfo['maxInputChannels'], input_format=pyaudio.paInt16): print 'Yay!' ... and it worked! So I am confused as to what the problem was. Any insight? – Bullheaded

Try upgrading portaudio, this fixed some rate problems for me. I used "brew install portaudio --HEAD". – Agatha

this worked for me, I did not realize soundcard's default sampling rate was 48khz, thanks! – Bloodstream

Thanks, this was my issue, too. I would not have expected this to be a problem on budget hardware but maybe 48k is becoming the de-facto norm? – Mazard

Answer 4

I worked this on OS X 10.10, Got the same error while trying to get audio from the microphone in a SYBA USB card (C Media chipset), and process it in real time with fft's and more:

IOError: [Errno Input overflowed] -9981

The overflow was completely solved when using a Callback Mode, instead of the Blocking Mode, as written by libbkmz.(https://www.python.org/dev/peps/pep-0263/)

Based on that, the bit of the working code looked like this:

"""
Creating the audio stream from our mic
"""
rate=48000
self.chunk=2**12
width = 2

p = pyaudio.PyAudio()

# callback function to stream audio, another thread.
def callback(in_data,frame_count, time_info, status):
    self.audio = numpy.fromstring(in_data,dtype=numpy.int16)
    return (self.audio, pyaudio.paContinue)

#create a pyaudio object
self.inStream = p.open(format = p.get_format_from_width(width, unsigned=False),
                       channels=1,
                       rate=rate,
                       input=True,
                       frames_per_buffer=self.chunk,
                       stream_callback = callback)

"""
Setting up the array that will handle the timeseries of audio data from our input
"""
self.audio = numpy.empty((self.buffersize),dtype="int16")

    self.inStream.start_stream()

while True:
  try:
    self.ANY_FUNCTION() #any function to run parallel to the audio thread, running forever, until ctrl+C is pressed. 

  except KeyboardInterrupt:

    self.inStream.stop_stream()
    self.inStream.close()
    p.terminate()
    print("* Killed Process")
    quit()

This code will create a callback function, then create a stream object, start it and then loop in any function. A separate thread streams audio, and that stream is closed when the main loop is stopped. self.audio is used in any function. I also had problems with the thread running forever if not terminated.

Since Pyaudio runs this stream in a separate thread, and this made the audio stream stable, the Blocking mode might have been saturating depending on the speed or timing of the rest of the processes in the script.

Note that the chunk size is 2^12, but smaller chunks work just as well. There are other parameters I considered and played around with to make sure they all made sense:

• Chunk size larger or smaller(no effect)
• Number and format of bits for the words in the buffer, signed 16 bit in this case.
• signedness of variables(tried with unsigned and got saturation patterns)
• Nature of mic input, and selection as default in the system, gain etc.

Hope that works for someone!

Hagiolatry

Answer 5

My other answer solved the problem in most cases. However sometimes the error still occurs.

That was the reason why I scrapped pyaudio and switched to pyalsaaudio. My Raspy now smoothly records any sound.

import alsaaudio   
import numpy as np
import array

# constants
CHANNELS    = 1
INFORMAT    = alsaaudio.PCM_FORMAT_FLOAT_LE
RATE        = 44100
FRAMESIZE   = 1024

# set up audio input
recorder=alsaaudio.PCM(type=alsaaudio.PCM_CAPTURE)
recorder.setchannels(CHANNELS)
recorder.setrate(RATE)
recorder.setformat(INFORMAT)
recorder.setperiodsize(FRAMESIZE)


buffer = array.array('f')
while <some condition>:
    buffer.fromstring(recorder.read()[1])

data = np.array(buffer, dtype='f')

Mothy

Really helpful, thank you! I used regular list instead of array.array, it is simpler nevertheless it works well for me, so the main change is pyaudio =>pyalsaaudio. Also, my microphone required PCM_FORMAT_S16_LE. – Music

Answer 6

FORMAT = pyaudio.paInt16

Make sure to set the correct format, my internal microphone was set to 24 Bit (see Audio-Midi-Setup application).

Charissecharita

Answer 7

Instead of

chunk = 1024

use:

chunk = 4096

It worked for me on a USB microphone.

Psychochemical

Answer 8

I had the same issue on the really slow raspberry pi, but I was able to solve it (for most cases) by using the faster array module for storing the data.

import array
import pyaudio 

FORMAT = pyaudio.paInt16
CHANNELS = 1
INPUT_CHANNEL=2
RATE = 48000
CHUNK = 512

p = pyaudio.PyAudio()
stream = p.open(format=FORMAT,
                channels=CHANNELS,
                rate=RATE,
                input=INPUT_CHANNEL,
                frames_per_buffer =CHUNK)

print("* recording")


try:
    data = array.array('h')
    for i in range(0, int(RATE / CHUNK * RECORD_SECONDS)):
        data.fromstring(stream.read(CHUNK))
finally:
    stream.stop_stream()
    stream.close()
    p.terminate()

print("* done recording")

The content of data is rather binary afterwards. But you can use numpy.array(data, dtype='i') to get a numpy array of intergers.

Mothy

Answer 9

For me this helped: https://mcmap.net/q/391530/-ioerror-errno-input-overflowed-9981

I used multiprocessing to write the file in parallel to recording audio. This is my code:

recordAudioSamples.py

import pyaudio
import wave
import datetime
import signal
import ftplib
import sys
import os

# configuration for assos_listen
import config

# run the audio capture and send sound sample processes
# in parallel
from multiprocessing import Process

# CONFIG
CHUNK = config.chunkSize
FORMAT = pyaudio.paInt16
CHANNELS = 1
RATE = config.samplingRate
RECORD_SECONDS = config.sampleLength

# HELPER FUNCTIONS

# write to ftp
def uploadFile(filename):

    print("start uploading file: " + filename)
    # connect to container
    ftp = ftplib.FTP(config.ftp_server_ip, config.username, config.password)

    # write file
    ftp.storbinary('STOR '+filename, open(filename, 'rb'))
    # close connection
    ftp.quit()
    print("finished uploading: " +filename)

# write to sd-card
def storeFile(filename,frames):

    print("start writing file: " + filename)
    wf = wave.open(filename, 'wb')
    wf.setnchannels(CHANNELS)
    wf.setsampwidth(p.get_sample_size(FORMAT))
    wf.setframerate(RATE)
    wf.writeframes(b''.join(frames))
    wf.close()
    print(filename + " written")

# abort the sampling process
def signal_handler(signal, frame):
    print('You pressed Ctrl+C!')

    # close stream and pyAudio
    stream.stop_stream()
    stream.close()
    p.terminate()

    sys.exit(0)

# MAIN FUNCTION
def recordAudio(p, stream):

    sampleNumber = 0
    while (True):
        print("*  recording")
        sampleNumber = sampleNumber +1

        frames = []
        startDateTimeStr = datetime.datetime.now().strftime("%Y_%m_%d_%I_%M_%S_%f")
        for i in range(0, int(RATE / CHUNK * RECORD_SECONDS)):
            data = stream.read(CHUNK)
            frames.append(data)

        fileName =  str(config.sensorID) + "_" + startDateTimeStr + ".wav"

        # create a store process to write the file in parallel
        storeProcess = Process(target=storeFile, args=(fileName,frames))
        storeProcess.start()

        if (config.upload == True):
            # since waiting for the upload to finish will take some time
            # and we do not want to have gaps in our sample
            # we start the upload process in parallel
            print("start uploading...")
            uploadProcess = Process(target=uploadFile, args=(fileName,))
            uploadProcess.start()

# ENTRYPOINT FROM CONSOLE
if __name__ == '__main__':

    p = pyaudio.PyAudio()
    stream = p.open(format=FORMAT,
                    channels=CHANNELS,
                    rate=RATE,
                    input=True,
                    frames_per_buffer=CHUNK)

    # directory to write and read files from
    os.chdir(config.storagePath)

    # abort by pressing C
    signal.signal(signal.SIGINT, signal_handler)
    print('\n\n--------------------------\npress Ctrl+C to stop the recording')

    # start recording
    recordAudio(p, stream)

config.py

### configuration file for assos_listen
# upload
upload = False

# config for this sensor
sensorID = "al_01"

# sampling rate & chunk size
chunkSize = 8192
samplingRate = 44100 # 44100 needed for Aves sampling
# choices=[4000, 8000, 16000, 32000, 44100] :: default 16000

# sample length in seconds
sampleLength = 10

# configuration for assos_store container
ftp_server_ip = "192.168.0.157"
username = "sensor"
password = "sensor"

# storage on assos_listen device
storagePath = "/home/pi/assos_listen_pi/storage/"

Paisley

Why not use a thread? Blocking I/O releases the GIL, making effective use of multiple cores without the complexities of multiprocessing. – Freedafreedman

Answer 10

This was helpful for me:

input_ = stream.read(chunk, exception_on_overflow=False)
exception_on_overflow = False

Jurisconsult