anon_sttts.yaml

data_dir: data
results_dir: wav # output example ./data/IEMOCAP_ims_sttts_pc/wav

pipeline: sttts
anon_suffix: !ref _<pipeline>

datasets:
  - name: IEMOCAP_dev
    data: IEMOCAP_dev
  - name: IEMOCAP_test
    data: IEMOCAP_test
  - name: libri_dev
    data: libri_dev
    enrolls: [_enrolls]
    trials: [_trials_f, _trials_m]
  - name: libri_test
    data: libri_test
    enrolls: [_enrolls]
    trials: [_trials_f, _trials_m]
  - name: train-clean-360
    data: train-clean-360

models_dir:  exp/sttts_models
save_intermediate: true
intermediate_dir: !ref exp/anon_pipeline_<pipeline>
# For faster inferance, download precomputed prosody/speaker_embedding/phn_transcript (libri only)
download_precomputed_intermediate_repr: true

modules:
  asr:
    recognizer: ims
    force_compute_recognition: false
    model_path: !ref <models_dir>/asr/asr_branchformer_tts-phn_en.zip
    ctc_weight: 0.2
    utt_start_token: "~"
    utt_end_token: "~#"
    results_path: !ref <intermediate_dir>/transcription/asr_branchformer_tts-phn_en

  speaker_embeddings:
    anonymizer: ims
    force_compute_extraction: false
    force_compute_anonymization: false
    vec_type: style-embed
    emb_model_path: !ref <models_dir>/tts/Embedding/embedding_function.pt
    anon_settings:
      method: gan
      vectors_file: !ref <models_dir>/anonymization/<modules[speaker_embeddings][vec_type]>_wgan.pt
      gan_model_path: !ref <models_dir>/anonymization/gan_<modules[speaker_embeddings][vec_type]>/<modules[speaker_embeddings][vec_type]>_wgan.pt
      num_sampled: 5000
      sim_threshold: 0.7
    extraction_results_path: !ref <intermediate_dir>/original_speaker_embeddings/<modules[speaker_embeddings][vec_type]>
    anon_results_path: !ref <intermediate_dir>/anon_speaker_embeddings/<modules[speaker_embeddings][vec_type]>
    anon_level_spk: []
    anon_level_utt: [IEMOCAP_test, IEMOCAP_dev, libri_dev, libri_test, train-clean-360]

  prosody:
    extractor_type: ims
    force_compute_extraction: false
    aligner_model_path: !ref <models_dir>/tts/Aligner/aligner.pt
    extraction_results_path: !ref <intermediate_dir>/original_prosody/ims_extractor
    anonymizer_type: ims
    random_offset_lower: 60
    random_offset_higher: 140
    anon_results_path: !ref <intermediate_dir>/anon_prosody/random_offsets

  tts:
    synthesizer: ims
    fastspeech_path: !ref <models_dir>/tts/FastSpeech2_Multi/prosody_cloning.pt
    hifigan_path: !ref <models_dir>/tts/HiFiGAN_combined/best.pt
    embeddings_path: !ref <models_dir>/tts/Embedding/embedding_function.pt
    output_sr: 16000
    results_path: !ref <intermediate_dir>/anon_speech/ims_sttts_pc