Add WAV/MP3 input with automatic 48 kHz resampling and stereo upmix

[Copilot]

The --src-audio (cover mode) and neural-codec --encode paths only accepted WAV at exactly 48 kHz. This adds transparent WAV + MP3 support at any sample rate, auto-resampled to 48 kHz and always delivered as stereo — exactly what the VAE encoder requires — with no ffmpeg pre-conversion needed.

New: src/audio.h

Single header providing read_audio(path, T_audio, n_channels):

• Format detected by extension: .mp3 → dr_mp3, anything else → dr_wav
• Linear resampler (audio_resample_linear) is channel-agnostic; only runs when sr ≠ 48000
• Always returns interleaved stereo [T × 2] — mono input is upmixed (L = R), N-channel input uses the first two channels; *n_channels is always 2 on success
• Uses free() to release dr_libs buffers (both dr_wav and dr_mp3 use the system allocator)
• Returns malloc'd buffer; caller frees

New: thirdparty/

• dr_wav.h v0.14.5 — WAV decode (public domain / MIT-0, mackron/dr_libs)
• dr_mp3.h v0.7.3 — MP3 decode via minimp3 (public domain / MIT-0)

Zero new link-time dependencies — both are single-header, included once per translation unit via #define DR_*_IMPLEMENTATION inside audio.h.

Tool changes

• neural-codec.cpp: encode path switches read_wav() → read_audio(); help text updated
• dit-vae.cpp: --src-audio switches to read_audio(); help text updated
• CMakeLists.txt: thirdparty/ added as SYSTEM include in the shared link_ggml_backends macro (vendor warnings suppressed)
• No existing source files (vae-enc.h etc.) were modified

Example

# Before: required exact 48 kHz stereo WAV, manual ffmpeg conversion otherwise
# After: any of these just work
./dit-vae --src-audio reference.mp3   ...
./dit-vae --src-audio reference.wav   ...  # any sample rate, mono or stereo
./neural-codec --vae vae.gguf --encode -i song.mp3 -o song.latent

New example

examples/cover.sh + examples/cover.json — demonstrates cover-mode generation from a WAV or MP3 reference track with inline usage notes.

---

🔒 GitHub Advanced Security automatically protects Copilot coding agent pull requests. You can protect all pull requests by enabling Advanced Security for your repositories. Learn more about Advanced Security.

[lmangani]

@copilot review the following report. remember we should convert stereo->stereo and mono->stereo and NOT touch the acestep-cpp existing files which work when invoked manually.

Known issues in src/audio.h (pending upstream fix)

Bug: mono audio is not upmixed to stereo before encoding

read_audio() returns a native-channel-count buffer ([T x n_channels] floats),
but vae_enc_compute() in vae-enc.h always reads two channels:

// vae-enc.h (hardcodes stereo access — UB when n_channels == 1)
for (int c = 0; c < 2; c++) {
    for (int t = 0; t < T_audio; t++) {
        m->scratch_in[c * T_audio + t] = audio[t * 2 + c];
    }
}

For stereo inputs (most user uploads) this works correctly.
For mono inputs the second channel index reads out-of-bounds memory.

Required fix in src/audio.h — always return interleaved stereo [T x 2].
Add this block after resampling completes (before the final return out):

// Upmix mono -> stereo, or use first two channels of N-ch audio.
if ((int) channels != 2) {
    int    n_ch_src = (int) channels;
    float *stereo   = (float *) malloc((size_t) T_raw * 2 * sizeof(float));
    if (!stereo) {
        fprintf(stderr, "[Audio] Out of memory converting to stereo\n");
        free(out);
        return NULL;
    }
    for (int t = 0; t < T_raw; t++) {
        float L = out[(size_t) t * n_ch_src + 0];
        float R = (n_ch_src > 1) ? out[(size_t) t * n_ch_src + 1] : L;
        stereo[t * 2 + 0] = L;
        stereo[t * 2 + 1] = R;
    }
    free(out);
    out = stereo;
    fprintf(stderr, "[Audio] Converted %dch -> stereo\n", n_ch_src);
}
*n_channels = 2;

Also: replace drwav_free(raw, NULL) with free(raw) in the resampling
branch — both dr_wav and dr_mp3 use the system allocator by default, so
free() is always safe regardless of which decoder produced the buffer.

[Copilot]

@copilot review the following report. remember we should convert stereo->stereo and mono->stereo and NOT touch the acestep-cpp existing files which work when invoked manually.

##...

Fixed in 7ea3602.

read_audio() now always returns interleaved stereo [T * 2] and sets *n_channels = 2:

• Mono → stereo: L = R = channel[0] (upmix after resampling, so both resampling and upmix work for mono at any rate)
• N > 2 channels → stereo: uses first two channels only
• drwav_free(raw, NULL) → free(raw): both dr_libs use the system allocator by default, so this is always correct

No existing files (vae-enc.h etc.) were touched.