Audio formats

WAV duration reported as nonsense in some readers

Streamed WAV responses are produced by an incremental encoder. Because the total size is not known until the stream ends, the RIFF size fields are not back-patched the way a fully buffered file’s would be, and a streamed WAV finalizes without rewriting them.

Most tools handle this fine (soundfile, pydub/ffmpeg, browsers, OS players). Python’s standard-library wave module does not, and may report a wrong duration or fail to read the length.

To get an exact duration, use a robust reader:

import soundfile as sf
print(sf.info("speech.wav").duration)

Or on the command line:

ffprobe -v error -show_entries format=duration -of default=nokey=1:noprint_wrappers=1 speech.wav

If you specifically need a WAV with fully correct header sizes, request it non-streamed ("stream": false), which returns a complete file, or transcode with ffmpeg.

Playing raw PCM

pcm output is raw signed 16-bit little-endian, 24 kHz, mono, with no header. Players need those parameters supplied explicitly:

ffplay -f s16le -ar 24000 -ch_layout mono speech.pcm

In code, wrap it in a WAV header or feed it to an audio device configured for paInt16 / 24000 Hz / mono. See Streaming and audio formats.

Choosing a format

Streaming to a player incrementally → pcm (simplest) or mp3.
A self-contained file → mp3, flac, or non-streamed wav.
Smallest lossy size → opus.
Lossless archival → flac.

aac is accepted by the request schema but is not currently produced; use one of the formats above.