Is this intended behavior?

[floorwarior]

i have a pipeline that "chunks" text using 'nltk', and randomly a "." made it into the text to audio queue, to my surprise this raises an error:

code:

voice = PiperVoice.load("pipermodels/en_GB-alan-medium.onnx")
with wave.open("test.wav", "wb") as wav_file:
    voice.synthesize_wav(". . . ", wav_file)

Traceback (most recent call last):
  File "c:\Users\ishall\Desktop\public_brainrootreader\brainrootreader\foolingaround.py", line 8, in <module>
    with wave.open("test.wav", "wb") as wav_file:
  File "C:\Users\ishall\AppData\Local\Programs\Python\Python312\Lib\wave.py", line 471, in __exit__
    self.close()
  File "C:\Users\ishall\AppData\Local\Programs\Python\Python312\Lib\wave.py", line 583, in close
    self._ensure_header_written(0)
  File "C:\Users\ishall\AppData\Local\Programs\Python\Python312\Lib\wave.py", line 601, in _ensure_header_written
    raise Error('# channels not specified')
wave.Error: # channels not specified

personally, i preferred if there is no audio being generated piper would still make a 0s audio file.
Is this intended behavior?

[yhay81]

This is an edge case in the current synthesize_wav implementation, rather than an intentional WAV representation.

voice.synthesize() skips empty phoneme lists, so punctuation-only input can produce zero AudioChunks (source). synthesize_wav() initializes the WAV format only when it receives the first chunk (source). With ". . . " there is no first chunk, so Python's wave writer reaches close() without nchannels ever being set and raises the error you saw.

If your desired contract is “always create a valid WAV, even when it has zero frames”, initialize the format yourself and tell Piper not to set it lazily:

import wave
from piper import PiperVoice

voice = PiperVoice.load("pipermodels/en_GB-alan-medium.onnx")

with wave.open("test.wav", "wb") as wav_file:
    wav_file.setframerate(voice.config.sample_rate)
    wav_file.setsampwidth(2)   # int16 PCM
    wav_file.setnchannels(1)   # Piper currently emits mono

    voice.synthesize_wav(
        ". . . ",
        wav_file,
        set_wav_format=False,
    )

For normal text this writes the audio as usual. For input that yields no chunks, it closes as a valid mono/16-bit WAV with nframes == 0 instead of throwing.

I would treat the exception as a small API edge-case/bug: the current AudioChunk format is already known (sample_width=2, sample_channels=1), so synthesize_wav could initialize those fields before entering the loop. Until that is changed upstream, setting the header explicitly is more robust than filtering on text.strip(), since punctuation-only text is non-empty but may still phonemize to no speakable content.