New variance parameter: voicing (#170)

* Support voicing in acoustic and variance models (experimental)

* Fix KeyError

configs/acoustic.yaml

@@ -54,12 +54,14 @@ mel_vmax: 1.5
 interp_uv: true
 energy_smooth_width: 0.12
 breathiness_smooth_width: 0.12
+voicing_smooth_width: 0.12
 tension_smooth_width: 0.12
 
 use_spk_id: false
 f0_embed_type: continuous
 use_energy_embed: false
 use_breathiness_embed: false
+use_voicing_embed: false
 use_tension_embed: false
 use_key_shift_embed: false
 use_speed_embed: false

configs/templates/config_acoustic.yaml

@@ -26,6 +26,8 @@ use_spk_id: false
 num_spk: 1
 use_energy_embed: false
 use_breathiness_embed: false
+use_voicing_embed: false
+use_tension_embed: false
 use_key_shift_embed: true
 use_speed_embed: true
 

configs/templates/config_variance.yaml

@@ -19,17 +19,26 @@ binary_data_dir: data/xxx/binary
 binarization_args:
   num_workers: 0
 
-energy_db_min: -96.0
-energy_db_max: -12.0
-breathiness_db_min: -96.0
-breathiness_db_max: -20.0
-
 use_spk_id: false
 num_spk: 1
 predict_dur: true
 predict_pitch: true
 predict_energy: false
 predict_breathiness: false
+predict_voicing: false
+predict_tension: false
+
+energy_db_min: -96.0
+energy_db_max: -12.0
+
+breathiness_db_min: -96.0
+breathiness_db_max: -20.0
+
+voicing_db_min: -96.0
+voicing_db_max: -12.0
+
+tension_logit_min: -10.0
+tension_logit_max: 10.0
 
 hidden_size: 256
 dur_prediction_args:

configs/variance.yaml

@@ -40,6 +40,7 @@ predict_dur: true
 predict_pitch: true
 predict_energy: false
 predict_breathiness: false
+predict_voicing: false
 predict_tension: false
 
 dur_prediction_args:
@@ -75,13 +76,22 @@ pitch_prediction_args:
 energy_db_min: -96.0
 energy_db_max: -12.0
 energy_smooth_width: 0.12
+
 breathiness_db_min: -96.0
 breathiness_db_max: -20.0
 breathiness_smooth_width: 0.12
 tension_logit_min: -10.0
 tension_logit_max: 10.0
 tension_smooth_width: 0.12
 
+voicing_db_min: -96.0
+voicing_db_max: -12.0
+voicing_smooth_width: 0.12
+
+tension_logit_min: -10.0
+tension_logit_max: 10.0
+tension_smooth_width: 0.12
+
 variances_prediction_args:
   total_repeat_bins: 48
   residual_layers: 10

inference/ds_acoustic.py

@@ -32,6 +32,8 @@ def __init__(self, device=None, load_model=True, load_vocoder=True, ckpt_steps=N
                 self.variances_to_embed.add('energy')
             if hparams.get('use_breathiness_embed', False):
                 self.variances_to_embed.add('breathiness')
+            if hparams.get('use_voicing_embed', False):
+                self.variances_to_embed.add('voicing')
             if hparams.get('use_tension_embed', False):
                 self.variances_to_embed.add('tension')
 

modules/fastspeech/acoustic_encoder.py

@@ -36,11 +36,14 @@ def __init__(self, vocab_size):
         self.variance_embed_list = []
         self.use_energy_embed = hparams.get('use_energy_embed', False)
         self.use_breathiness_embed = hparams.get('use_breathiness_embed', False)
+        self.use_voicing_embed = hparams.get('use_voicing_embed', False)
         self.use_tension_embed = hparams.get('use_tension_embed', False)
         if self.use_energy_embed:
             self.variance_embed_list.append('energy')
         if self.use_breathiness_embed:
             self.variance_embed_list.append('breathiness')
+        if self.use_voicing_embed:
+            self.variance_embed_list.append('voicing')
         if self.use_tension_embed:
             self.variance_embed_list.append('tension')
 

modules/fastspeech/param_adaptor.py

@@ -5,7 +5,7 @@
 from modules.diffusion.ddpm import MultiVarianceDiffusion
 from utils.hparams import hparams
 
-VARIANCE_CHECKLIST = ['energy', 'breathiness', 'tension']
+VARIANCE_CHECKLIST = ['energy', 'breathiness', 'voicing', 'tension']
 
 
 class ParameterAdaptorModule(torch.nn.Module):
@@ -14,11 +14,14 @@ def __init__(self):
         self.variance_prediction_list = []
         self.predict_energy = hparams.get('predict_energy', False)
         self.predict_breathiness = hparams.get('predict_breathiness', False)
+        self.predict_voicing = hparams.get('predict_voicing', False)
         self.predict_tension = hparams.get('predict_tension', False)
         if self.predict_energy:
             self.variance_prediction_list.append('energy')
         if self.predict_breathiness:
             self.variance_prediction_list.append('breathiness')
+        if self.predict_voicing:
+            self.variance_prediction_list.append('voicing')
         if self.predict_tension:
             self.variance_prediction_list.append('tension')
         self.predict_variances = len(self.variance_prediction_list) > 0
@@ -41,6 +44,13 @@ def build_adaptor(self, cls=MultiVarianceDiffusion):
             ))
             clamps.append((hparams['breathiness_db_min'], 0.))
 
+        if self.predict_voicing:
+            ranges.append((
+                hparams['voicing_db_min'],
+                hparams['voicing_db_max']
+            ))
+            clamps.append((hparams['voicing_db_min'], 0.))
+
         if self.predict_tension:
             ranges.append((
                 hparams['tension_logit_min'],

preprocessing/acoustic_binarizer.py

@@ -26,6 +26,7 @@
     get_mel2ph_torch,
     get_energy_librosa,
     get_breathiness_pyworld,
+    get_voicing_pyworld,
     get_tension_base_harmonic,
 )
 from utils.hparams import hparams
@@ -39,6 +40,7 @@
     'f0',
     'energy',
     'breathiness',
+    'voicing',
     'tension',
     'key_shift',
     'speed',
@@ -47,6 +49,7 @@
 pitch_extractor: BasePE = None
 energy_smooth: SinusoidalSmoothingConv1d = None
 breathiness_smooth: SinusoidalSmoothingConv1d = None
+voicing_smooth: SinusoidalSmoothingConv1d = None
 tension_smooth: SinusoidalSmoothingConv1d = None
 
 
@@ -56,6 +59,7 @@ def __init__(self):
         self.lr = LengthRegulator()
         self.need_energy = hparams['use_energy_embed']
         self.need_breathiness = hparams['use_breathiness_embed']
+        self.need_voicing = hparams['use_voicing_embed']
         self.need_tension = hparams['use_tension_embed']
 
     def load_meta_data(self, raw_data_dir: pathlib.Path, ds_id, spk_id):
@@ -158,6 +162,21 @@ def process_item(self, item_name, meta_data, binarization_args):
 
             processed_input['breathiness'] = breathiness.cpu().numpy()
 
+        if self.need_voicing:
+            # get ground truth voicing
+            voicing = get_voicing_pyworld(
+                dec_waveform, None, None, length=length
+            )
+
+            global voicing_smooth
+            if voicing_smooth is None:
+                voicing_smooth = SinusoidalSmoothingConv1d(
+                    round(hparams['voicing_smooth_width'] / self.timestep)
+                ).eval().to(self.device)
+            voicing = voicing_smooth(torch.from_numpy(voicing).to(self.device)[None])[0]
+
+            processed_input['voicing'] = voicing.cpu().numpy()
+
         if self.need_tension:
             # get ground truth tension
             tension = get_tension_base_harmonic(

preprocessing/variance_binarizer.py

@@ -19,6 +19,7 @@
     get_mel2ph_torch,
     get_energy_librosa,
     get_breathiness_pyworld,
+    get_voicing_pyworld,
     get_tension_base_harmonic,
 )
 from utils.hparams import hparams
@@ -44,6 +45,7 @@
     'uv',  # unvoiced masks (only for objective evaluation metrics), bool[T_s,]
     'energy',  # frame-level RMS (dB), float32[T_s,]
     'breathiness',  # frame-level RMS of aperiodic parts (dB), float32[T_s,]
+    'voicing',  # frame-level RMS of harmonic parts (dB), float32[T_s,]
     'tension',  # frame-level tension (logit), float32[T_s,]
 ]
 DS_INDEX_SEP = '#'
@@ -54,6 +56,7 @@
 midi_smooth: SinusoidalSmoothingConv1d = None
 energy_smooth: SinusoidalSmoothingConv1d = None
 breathiness_smooth: SinusoidalSmoothingConv1d = None
+voicing_smooth: SinusoidalSmoothingConv1d = None
 tension_smooth: SinusoidalSmoothingConv1d = None
 
 
@@ -74,8 +77,9 @@ def __init__(self):
 
         predict_energy = hparams['predict_energy']
         predict_breathiness = hparams['predict_breathiness']
+        predict_voicing = hparams['predict_voicing']
         predict_tension = hparams['predict_tension']
-        self.predict_variances = predict_energy or predict_breathiness or predict_tension
+        self.predict_variances = predict_energy or predict_breathiness or predict_voicing or predict_tension
         self.lr = LengthRegulator().to(self.device)
         self.prefer_ds = self.binarization_args['prefer_ds']
         self.cached_ds = {}
@@ -417,6 +421,37 @@ def process_item(self, item_name, meta_data, binarization_args):
 
             processed_input['breathiness'] = breathiness
 
+        # Below: extract voicing
+        if hparams['predict_voicing']:
+            voicing = None
+            voicing_from_wav = False
+            if self.prefer_ds:
+                voicing_seq = self.load_attr_from_ds(ds_id, name, 'voicing', idx=ds_seg_idx)
+                if voicing_seq is not None:
+                    voicing = resample_align_curve(
+                        np.array(voicing_seq.split(), np.float32),
+                        original_timestep=float(self.load_attr_from_ds(
+                            ds_id, name, 'voicing_timestep', idx=ds_seg_idx
+                        )),
+                        target_timestep=self.timestep,
+                        align_length=length
+                    )
+            if voicing is None:
+                voicing = get_voicing_pyworld(
+                    dec_waveform, None, None, length=length
+                )
+                voicing_from_wav = True
+
+            if voicing_from_wav:
+                global voicing_smooth
+                if voicing_smooth is None:
+                    voicing_smooth = SinusoidalSmoothingConv1d(
+                        round(hparams['voicing_smooth_width'] / self.timestep)
+                    ).eval().to(self.device)
+                voicing = voicing_smooth(torch.from_numpy(voicing).to(self.device)[None])[0].cpu().numpy()
+
+            processed_input['voicing'] = voicing
+
         # Below: extract tension
         if hparams['predict_tension']:
             tension = None

training/acoustic_task.py

@@ -27,6 +27,8 @@ def __init__(self, prefix, preload=False):
             self.required_variances['energy'] = 0.0
         if hparams['use_breathiness_embed']:
             self.required_variances['breathiness'] = 0.0
+        if hparams['use_voicing_embed']:
+            self.required_variances['voicing'] = 0.0
         if hparams['use_tension_embed']:
             self.required_variances['tension'] = 0.0
 
@@ -80,6 +82,8 @@ def __init__(self):
             self.required_variances.append('energy')
         if hparams['use_breathiness_embed']:
             self.required_variances.append('breathiness')
+        if hparams['use_voicing_embed']:
+            self.required_variances.append('voicing')
         if hparams['use_tension_embed']:
             self.required_variances.append('tension')
         super()._finish_init()

training/variance_task.py

@@ -24,8 +24,9 @@ def __init__(self, prefix, preload=False):
         super(VarianceDataset, self).__init__(prefix, hparams['dataset_size_key'], preload)
         need_energy = hparams['predict_energy']
         need_breathiness = hparams['predict_breathiness']
+        need_voicing = hparams['predict_voicing']
         need_tension = hparams['predict_tension']
-        self.predict_variances = need_energy or need_breathiness or need_tension
+        self.predict_variances = need_energy or need_breathiness or need_voicing or need_tension
 
     def collater(self, samples):
         batch = super().collater(samples)
@@ -60,6 +61,8 @@ def collater(self, samples):
             batch['energy'] = utils.collate_nd([s['energy'] for s in samples], 0)
         if hparams['predict_breathiness']:
             batch['breathiness'] = utils.collate_nd([s['breathiness'] for s in samples], 0)
+        if hparams['predict_voicing']:
+            batch['voicing'] = utils.collate_nd([s['voicing'] for s in samples], 0)
         if hparams['predict_tension']:
             batch['tension'] = utils.collate_nd([s['tension'] for s in samples], 0)
 
@@ -92,12 +95,15 @@ def __init__(self):
 
         predict_energy = hparams['predict_energy']
         predict_breathiness = hparams['predict_breathiness']
+        predict_voicing = hparams['predict_voicing']
         predict_tension = hparams['predict_tension']
         self.variance_prediction_list = []
         if predict_energy:
             self.variance_prediction_list.append('energy')
         if predict_breathiness:
             self.variance_prediction_list.append('breathiness')
+        if predict_voicing:
+            self.variance_prediction_list.append('voicing')
         if predict_tension:
             self.variance_prediction_list.append('tension')
         self.predict_variances = len(self.variance_prediction_list) > 0
@@ -153,6 +159,7 @@ def run_model(self, sample, infer=False):
         pitch = sample.get('pitch')  # [B, T_s]
         energy = sample.get('energy')  # [B, T_s]
         breathiness = sample.get('breathiness')  # [B, T_s]
+        voicing = sample.get('voicing')  # [B, T_s]
         tension = sample.get('tension')  # [B, T_s]
 
         pitch_retake = variance_retake = None
@@ -175,7 +182,7 @@ def run_model(self, sample, infer=False):
             note_midi=note_midi, note_rest=note_rest,
             note_dur=note_dur, note_glide=note_glide, mel2note=mel2note,
             base_pitch=base_pitch, pitch=pitch,
-            energy=energy, breathiness=breathiness, tension=tension,
+            energy=energy, breathiness=breathiness, voicing=voicing, tension=tension,
             pitch_retake=pitch_retake, variance_retake=variance_retake,
             spk_id=spk_ids, infer=infer
         )

utils/binarizer_utils.py

@@ -279,6 +279,35 @@ def get_breathiness_pyworld(
     return breathiness
 
 
+def get_voicing_pyworld(
+        waveform: Union[np.ndarray, DecomposedWaveform],
+        samplerate, f0, length,
+        *, hop_size=None, fft_size=None, win_size=None
+):
+    """
+    Definition of voicing: RMS of the harmonic part, in dB representation
+    :param waveform: All other analysis parameters will not take effect if a DeconstructedWaveform is given
+    :param samplerate: sampling rate
+    :param f0: reference f0
+    :param length: Expected number of frames
+    :param hop_size: Frame width, in number of samples
+    :param fft_size: Number of fft bins
+    :param win_size: Window size, in number of samples
+    :return: voicing
+    """
+    if not isinstance(waveform, DecomposedWaveform):
+        waveform = DecomposedWaveform(
+            waveform=waveform, samplerate=samplerate, f0=f0,
+            hop_size=hop_size, fft_size=fft_size, win_size=win_size
+        )
+    waveform_sp = waveform.harmonic()
+    voicing = get_energy_librosa(
+        waveform_sp, length=length,
+        hop_size=waveform.hop_size, win_size=waveform.win_size
+    )
+    return voicing
+
+
 def get_tension_base_harmonic(
         waveform: Union[np.ndarray, DecomposedWaveform],
         samplerate, f0, length,