Audio upload and conversion

Use Convert to change formats without loading or running an inference model. All four workflows share the same output-format selector, downloads and ZIP manifests. Conversion jobs use the existing queue, authentication, upload/duration limits, deletion and expiry rules.

Download	Encoding	Sample rate
wav	Float32 PCM	Workflow rate
flac	24-bit PCM	Workflow rate
opus	Opus, 96 kbps VBR	48 kHz
mp3	MP3, 192 kbps	48 kHz
m4a	AAC, 192 kbps, MPEG-4 container	48 kHz
aac	AAC, 192 kbps, ADTS container	48 kHz
ogg	Vorbis, quality 5	48 kHz

Conversion keeps channels, with MP3 downmixed to at most stereo. WAV/FLAC retain the input sample rate. Enhancement resynthesizes mono at 48 kHz; separation produces two mono tracks at 16 kHz before export. Compressed export resamples each track to 48 kHz. Export does not apply loudness normalization. Float WAV preserves peaks above full scale; FLAC clips them. Compressed encodings are lossy and may add codec padding or slightly change decoded duration/peaks. Manifest duration describes the source timeline before encoding.

Uploads are inspected by content, not filename. libsndfile handles its supported recordings; the FFmpeg fallback adds AAC/M4A, WebM audio and other allowed audio containers. Playlists and network protocols are excluded. Decoding limits remain 8–192 kHz, at most eight channels and the configured duration/upload size. Enhancement/separation downmix input; Convert preserves its channels. Browser playback support varies by codec; files can still be converted and downloaded.

curl -X POST http://localhost:7868/api/jobs -F file=@voice.m4a -F mode=convert -F output_format=opus
curl http://localhost:7868/api/jobs/JOB_ID
curl -OJ http://localhost:7868/api/jobs/JOB_ID/files/converted.opus

Docker includes FFmpeg/ffprobe, codec notices and matching Debian sources in the source image. Native Python users must install FFmpeg and ffprobe on PATH. WAV and FLAC do not require FFmpeg. No Python model dependencies were added.