README.md

MOLM-Audio: SPDMark-Style Segment-Wise Audio Watermarking

LoRA-routing audio watermarks for three generators (HiFi-GAN, VibeVoice acoustic decoder, DiffWave). Per SPDMark (https://arxiv.org/abs/2512.12090), each audio is split into S=8 segments; each segment carries an HMAC-derived M-bit message embedded via a parallel LoRA "basis dictionary." The verifier matches recovered per-segment bits against the expected HMAC sequence using Hungarian assignment + Binomial hypothesis test.

Repo layout

checkpoints/
  diffwave_spdmark_spec/final/{lora_weights.pt, extractor.pt, diffwave_full.pt}
  diffwave_v5_2step/final/{lora_weights.pt, extractor.pt, diffwave_full.pt}
  hifigan_spdmark_spec/final/{lora_weights.pt, extractor.pt, hifigan_full.pt}
  vibevoice_spdmark_spec/final/{lora_weights.pt, extractor.pt, model_full.pt}
  vibevoice_14bit_v2/final/{lora_weights.pt, extractor.pt, model_full.pt}
code/                  # training + inference + smoke-test scripts (clone or download)
eval/<run>_<regime>/results.json
README.md

lora_weights.pt + the base model from torchaudio/pretrained is sufficient for inference. *_full.pt is provided for one-step loading where the base model isn't available locally.

Training/test wavs live in MOLM-Audio/molm-audio-data under data/.

Setup

# 1. Clone or download this repo.
hf download MOLM-Audio/molm-audio --local-dir molm-audio

# 2. Drop the code/ contents into a Python env.
cd molm-audio/code
pip install -r requirements.txt

# 3. For DiffWave: also place the LJSpeech base checkpoint at
#    pretrained/diffwave-ljspeech.pt (download from the upstream DiffWave repo).
# 4. For VibeVoice: extract decoder + diffusion head components once via:
python precompute_vibevoice_data.py extract \
    --vibevoice_model microsoft/VibeVoice-1.5B \
    --output_dir pretrained/vibevoice
# Then encode your audio into latents with `precompute_vibevoice_data.py encode`.
# 5. For HiFi-GAN: nothing extra — torchaudio downloads V3 LJSpeech weights on first run.

Checkpoints

Current runs (segments, spec-trained, with Hungarian verifier)

Run	Backbone	Routing	Paths	Bits/seg	Train attack
hifigan_spdmark_spec	HiFi-GAN V3 (torchaudio LJSpeech)	0,1,2,3,4,5,6	4	14	nvlceqr
vibevoice_spdmark_spec	VibeVoice acoustic decoder	0,1,2,3,4,5,6 (×2 slots)	2	14	nvlceqr
diffwave_spdmark_spec	DiffWave (LJSpeech)	0,4,8,12,16,20,24	4	14	nvlceqr

Legacy runs (no segments)

Run	Backbone	Routing	Paths	Bits/seg	Train attack	Notes
vibevoice_14bit_v2	VibeVoice acoustic decoder	0,1,2,3,4,5,6 (×2 slots)	2	14	nvlceq	Pre-SPDMark-temporal-attacks training; eval JSON has no Hungarian verify block.
diffwave_v5_2step	DiffWave (LJSpeech)	0,4,8,12,16,20,24	2	7	nvlceq	2-step diffusion, lambda_perc 0.1 → louder watermark, faster inference.

Spectral attack codes: n noise, v gain, l lowpass, c crop, e erase, q quantize, r resample. SPDMark temporal codes: d segment-drop, s segment-swap, i segment-insert.

Training

All three trainers share the SPDMark plumbing in code/molm_audio_adapter.py (HMAC keys via sample_training_keys, routing-mask construction, segment-map), code/audio_augmentations.py (segment-aware aug with optional temporal attacks), and code/audio_losses.py (SegmentCombinedAudioLoss with valid_mask).

The exact commands used to produce the spec-trained checkpoints in this repo:

HiFi-GAN segments

python code/training_molm_hifigan.py \
    --dataset_path data/train_full \
    --output_dir checkpoints/hifigan_spdmark_spec \
    --exp_name MOLM_HiFiGAN_SPDMark_spec \
    --routing_blocks 0,1,2,3,4,5,6 --num_paths 4 --lora_rank 64 \
    --num_segments 8 --key_bits 128 \
    --attack nvlceqr --aug_prob 0.25 \
    --lambda_perc 0.5 --lambda_perc_warmup_steps 4500 \
    --max_train_steps 75000 --train_batch_size 8 --gradient_accumulation_steps 4 \
    --learning_rate 2e-4 --lora_init_std 0.07 \
    --audio_seconds 2.0 \
    --checkpointing_steps 500 --logging_steps 10 --use_wandb

DiffWave segments

python code/training_molm_audio.py \
    --diffwave_checkpoint pretrained/diffwave-ljspeech.pt \
    --dataset_path data/train_full \
    --output_dir checkpoints/diffwave_spdmark_spec \
    --exp_name DiffWave_SPDMark_spec \
    --routing_layers 0,4,8,12,16,20,24 --num_paths 4 --lora_rank 64 \
    --lora_alpha 24 --lora_init_std 0.01 \
    --diffusion_steps 4 \
    --num_segments 8 --key_bits 128 \
    --attack nvlceqr --aug_prob 0.25 \
    --adam_weight_decay 0.05 --lambda_lora_reg 0.01 \
    --lambda_perc 0.3 --lambda_perc_warmup_steps 2000 \
    --max_train_steps 40000 --train_batch_size 8 --gradient_accumulation_steps 2 \
    --learning_rate 3e-4 --max_grad_norm 1.0 --audio_seconds 2.0 \
    --checkpointing_steps 500 --logging_steps 10 --use_wandb

VibeVoice spec

python code/training_molm_vibevoice.py \
    --components_dir pretrained/vibevoice \
    --data_dir data/vibevoice_latents_train_full \
    --output_dir checkpoints/vibevoice_spdmark_spec \
    --exp_name vibevoice_spdmark_spec \
    --routing_blocks 0,1,2,3,4,5,6 \
    --num_paths 2 --lora_rank 16 --lora_alpha 8 \
    --num_frames 8 --num_segments 8 --key_bits 128 \
    --attack nvlceqr --aug_prob 0.3 \
    --lambda_perc 0.5 --lambda_perc_warmup_steps 4500 --lambda_lora_reg 0.01 \
    --max_train_steps 120000 --cosine_cycle 1 \
    --train_batch_size 16 --gradient_accumulation_steps 2 \
    --learning_rate 1e-4 \
    --checkpointing_steps 500 --logging_steps 10 --use_wandb

No segments training

diffwave_v5_2step (2-step diffusion, very low lambda_perc):

python code/training_molm_audio.py \
    --diffwave_checkpoint pretrained/diffwave-ljspeech.pt \
    --dataset_path data/train_full \
    --routing_layers 0,4,8,12,16,20,24 --num_paths 2 --lora_rank 64 --lora_alpha 64 \
    --diffusion_steps 2 \
    --max_train_steps 120000 --cosine_cycle 1 \
    --train_batch_size 8 --gradient_accumulation_steps 4 \
    --learning_rate 1e-4 \
    --attack nvlceq --aug_prob 0.3 \
    --lambda_perc 0.1 --lambda_perc_warmup_steps 4500 --lambda_lora_reg 0.01 \
    --output_dir checkpoints/diffwave_v5_2step \
    --checkpointing_steps 100 --logging_steps 10 --use_wandb \
    --exp_name diffwave_v5_2step

vibevoice_14bit_v2:

python code/training_molm_vibevoice.py \
    --components_dir pretrained/vibevoice \
    --data_dir data/vibevoice_latents_train_full \
    --routing_blocks 0,1,2,3,4,5,6 \
    --num_paths 2 --lora_rank 16 --lora_alpha 8 \
    --max_train_steps 120000 --cosine_cycle 1 \
    --train_batch_size 16 --gradient_accumulation_steps 2 \
    --learning_rate 1e-4 \
    --attack nvlceq --aug_prob 0.3 \
    --lambda_perc 0.5 --lambda_perc_warmup_steps 4500 --lambda_lora_reg 0.01 \
    --output_dir checkpoints/vibevoice_14bit_v2 \
    --exp_name vibevoice_14bit_v2 \
    --checkpointing_steps 500 --logging_steps 10 --use_wandb

Key SPDMark training flags (all three trainers):

• --num_segments 8 — S, segments per audio clip.
• --key_bits 128 — base-key width fed into HMAC-SHA256 for per-segment message derivation.
• --attack <codes> — see attack-code legend above. Append dsi to also train against segment-level temporal attacks (drop/swap/insert).

Inference

All three generators share a common eval interface; --verify enables the Hungarian + Binomial verifier (auto-sets --message_scheme hmac).

HiFi-GAN

python code/generate_molm_hifigan.py \
    --lora_weights checkpoints/hifigan_spdmark_spec/final/lora_weights.pt \
    --extractor_weights checkpoints/hifigan_spdmark_spec/final/extractor.pt \
    --routing_blocks 0,1,2,3,4,5,6 --num_paths 4 --lora_rank 64 \
    --test_dir <path-to-wavs> --num_samples 20 \
    --chunked_generation --num_chunks 8 \
    --verify --gamma_f 0.01 --gamma_v 0.01 --attacks nvlceqr \
    --output_dir eval_out/hifigan_spec_spectral --device cuda

DiffWave (current spdmark_spec)

python code/generate_molm_audio.py \
    --diffwave_checkpoint pretrained/diffwave-ljspeech.pt \
    --lora_weights checkpoints/diffwave_spdmark_spec/final/lora_weights.pt \
    --extractor_weights checkpoints/diffwave_spdmark_spec/final/extractor.pt \
    --routing_layers 0,4,8,12,16,20,24 --num_paths 4 --lora_rank 64 \
    --diffusion_steps 4 \
    --test_dir <path-to-wavs> --num_samples 20 \
    --chunked_generation --num_chunks 8 \
    --verify --gamma_f 0.01 --gamma_v 0.01 --attacks nvlceqr \
    --output_dir eval_out/diffwave_spec_spectral --device cuda

DiffWave (no segments v5_2step)

python code/generate_molm_audio.py \
    --diffwave_checkpoint pretrained/diffwave-ljspeech.pt \
    --lora_weights checkpoints/diffwave_v5_2step/final/lora_weights.pt \
    --extractor_weights checkpoints/diffwave_v5_2step/final/extractor.pt \
    --routing_layers 0,4,8,12,16,20,24 --num_paths 2 --lora_rank 64 \
    --diffusion_steps 2 \
    --test_dir <path-to-wavs> --num_samples 20 \
    --chunked_generation --num_chunks 8 \
    --verify --gamma_f 0.01 --gamma_v 0.01 --attacks nvlceqr \
    --output_dir eval_out/diffwave_v5_2step_spectral --device cuda

VibeVoice (current spdmark_spec)

python code/generate_molm_vibevoice.py \
    --components_dir pretrained/vibevoice \
    --lora_weights checkpoints/vibevoice_spdmark_spec/final/lora_weights.pt \
    --extractor_weights checkpoints/vibevoice_spdmark_spec/final/extractor.pt \
    --routing_blocks 0,1,2,3,4,5,6 --num_paths 2 --lora_rank 16 --lora_alpha 8 \
    --num_frames 8 \
    --data_dir <path-to-precomputed-latents> --num_samples 20 \
    --chunked_generation --num_chunks 8 \
    --verify --gamma_f 0.01 --gamma_v 0.01 --attacks nvlceqr \
    --output_dir eval_out/vibevoice_spec_spectral --device cuda

VibeVoice (no segments 14bit_v2)

Same flags as spdmark_spec but point at checkpoints/vibevoice_14bit_v2/final/.

Smoke tests

Quick sanity check of the SPDMark plumbing:

python code/smoke_test_hifigan.py    # 9 checks incl. HMAC keys + segment-aware aug + verify
python code/smoke_test.py            # DiffWave equivalents
python code/smoke_test_vibevoice.py  # VibeVoice equivalents