μ-law voice quality musings

[jozsefszalma]

I really like the PyVoIP library as it offers the right level of abstraction. I searched all over the net for a solution to my issue, so before anyone else spends days debugging audio quality problems with μ-law encoding, here are a couple of observations:

PyVoIP handles audio encoding and decoding as follows:

def encode_pcmu(self, packet: bytes) -> bytes:
    packet = audioop.bias(packet, 1, -128)  # Converts from unsigned to signed
    packet = audioop.lin2ulaw(packet, 1)    # Converts to μ-law
    return packet
  
def parse_pcmu(self, packet: RTPMessage) -> None:
    data = audioop.ulaw2lin(packet.payload, 1)  # Converts from μ-law to linear
    data = audioop.bias(data, 1, 128)           # Converts from signed to unsigned
    self.pmin.write(packet.timestamp, data)

This code assumes that you're passing 8-bit unsigned audio to call.write_audio, and the lin2ulaw conversion happens internally.

Implications:

• If you're new to VoIP (like me) and assume your endpoint requires μ-law encoded audio, thus you pre-process your audio yourself then that might not work out so well without knowing what happens under the hood.
• μ-law was designed to compress 16-bit audio down to 8-bits (while retaining 14 bits of dynamic range). Passing it 8-bit data can degrade audio quality.

To solve my own use-case I turned the encode_pcmu and the parse_pcmu functions in a fork to pass-through and deal with the encoding and decoding outside PyVoIP.