voice_audio_multi/docs/models/apollo.md

Apollo

Apollo restores compressed music at 44.1 kHz.

Quick Start

audiocpp_cli \
  --task s2s \
  --family apollo \
  --model models/Apollo-GGUF/apollo-orig.gguf \
  --backend cuda \
  --audio input_44100.wav \
  --out restored.wav \
  --log

Model

Field	Value
Family	apollo
Task	s2s
Mode	offline
Default weights	models/Apollo-GGUF/apollo-orig.gguf (F32)
Input	44.1 kHz WAV through --audio
Output	Restored 44.1 kHz waveform, preserving input channels

Common Options (use directly)

Option	Values	Default	Meaning
--audio	44.1 kHz WAV path	required	Music to restore.
--out	WAV path	not set	Save the restored waveform.

Request Options (use with --request-option)

Option	Values	Default	Meaning
audio_chunk_duration_sec	seconds >= 0	0	Chunk duration; 0 processes the whole recording. When enabled, the rounded chunk sample count must exceed n_fft/2.
audio_chunk_overlap_sec	seconds >= 0	1	Crossfade overlap; when chunking is enabled, at most half the chunk length after rounding to samples. Ignored for whole-file inference.
edge_pad_duration_sec	seconds >= 0	0	Extra context processed and discarded on each side of a chunk.

For long recordings, add these options to limit per-chunk memory use:

--request-option audio_chunk_duration_sec=6 \
--request-option audio_chunk_overlap_sec=1 \
--request-option edge_pad_duration_sec=1

Session Options (use with --session-option)

Option	Values	Default	Meaning
apollo.weight_type	native, f32, f16, bf16, q8_0, q4_0, q4_1, q5_0, q5_1, q2_k, q3_k, q4_k, q5_k, q6_k	native	Weight storage override at session creation.

Convert Weights

From the repository root, convert the official Apollo pytorch_model.bin:

python tests/apollo/convert_apollo_gguf.py \
  --checkpoint /path/to/pytorch_model.bin \
  --output models/Apollo-GGUF/apollo-orig.gguf \
  --log apollo-conversion.log

The script uses build/debug/bin/audiocpp_gguf by default; use --audiocpp-gguf /path/to/audiocpp_gguf for another executable.