transformers

bb8f2353 - Fix ASR pipeline mono conversion for channels-last audio (fixes #47886) (#47888)

Fix ASR pipeline mono conversion for channels-last audio (fixes #47886) (#47888) * Fix ASR pipeline mono conversion for channels-last audio The pipeline averaged multi-channel input over a hardcoded axis 0, which assumes channels-first. soundfile.read, librosa.load(mono=False) and scipy.io.wavfile.read all return channels-last, so a stereo waveform was averaged across time and collapsed to one value per channel: 48000 samples became 2, were padded back to silence, and transcribed as " you" with only a warning saying the conversion had succeeded. Infer the channel axis from the shorter dimension instead, and report the actual shape rather than ndim, which always printed 2 for any 2-D input. * Reject ambiguous channel layouts instead of inferring the channel axis Per review: inferring the channel axis with argmin is unsafe. It picks an arbitrary axis for ndim >= 3, and it silently downmixes 5.1 audio where the weighting should be the caller's choice. Keep the documented (channels, samples) layout, which is what torchcodec returns, and raise when the input does not conform. A channels-last array gets an error naming the transpose, rather than a transcription of two samples. Also move the conversion ahead of F.resample. The resampler operates on the last axis, so a multi-channel array had to be reduced first, and the stride arithmetic reads shape[0] as samples. * Correct the comment on why the downmix runs before resampling * Apply suggestion from @Rocketknight1 --------- Co-authored-by: Matt <Rocketknight1@users.noreply.github.com>