support presets for generation

api.py

@@ -160,6 +160,21 @@ class TextToSpeech:
         self.vocoder.load_state_dict(torch.load('.models/vocoder.pth')['model_g'])
         self.vocoder.eval(inference=True)
 
+    def tts_with_preset(self, text, voice_samples, preset='intelligible', **kwargs):
+        """
+        Calls TTS with one of a set of preset generation parameters. Options:
+            'intelligible': Maximizes the probability of understandable words at the cost of diverse voices, intonation and prosody.
+            'realistic': Increases the diversity of spoken voices and improves realism of vocal characteristics at the cost of intelligibility.
+            'mid': Somewhere between 'intelligible' and 'realistic'.
+        """
+        presets = {
+            'intelligible': {'temperature': .5, 'length_penalty': 2.0, 'repetition_penalty': 2.0, 'top_p': .5, 'diffusion_iterations': 100, 'cond_free': True, 'cond_free_k': .7, 'diffusion_temperature': .7},
+            'mid': {'temperature': .7, 'length_penalty': 1.0, 'repetition_penalty': 2.0, 'top_p': .7, 'diffusion_iterations': 100, 'cond_free': True, 'cond_free_k': 1.5, 'diffusion_temperature': .8},
+            'realistic': {'temperature': .9, 'length_penalty': 1.0, 'repetition_penalty': 1.3, 'top_p': .9, 'diffusion_iterations': 100, 'cond_free': True, 'cond_free_k': 2, 'diffusion_temperature': 1},
+        }
+        kwargs.update(presets[preset])
+        return self.tts(text, voice_samples, **kwargs)
+
     def tts(self, text, voice_samples, k=1,
             # autoregressive generation parameters follow
             num_autoregressive_samples=512, temperature=.5, length_penalty=1, repetition_penalty=2.0, top_p=.5,

eval_multiple.py

@@ -7,7 +7,7 @@ from utils.audio import load_audio
 
 if __name__ == '__main__':
     fname = 'Y:\\libritts\\test-clean\\transcribed-brief-w2v.tsv'
-    outpath = 'D:\\tmp\\tortoise-tts-eval\\compare_vocoders'
+    outpath = 'D:\\tmp\\tortoise-tts-eval\\eval_new_autoregressive'
     outpath_real = 'D:\\tmp\\tortoise-tts-eval\\real'
 
     os.makedirs(outpath, exist_ok=True)
@@ -24,16 +24,12 @@ if __name__ == '__main__':
         path = os.path.join(os.path.dirname(fname), line[1])
         cond_audio = load_audio(path, 22050)
         torchaudio.save(os.path.join(outpath_real, os.path.basename(line[1])), cond_audio, 22050)
-        sample, sample2 = tts.tts(transcript, [cond_audio, cond_audio], num_autoregressive_samples=512, k=1,
+        sample = tts.tts(transcript, [cond_audio, cond_audio], num_autoregressive_samples=512, k=1,
                              repetition_penalty=2.0, length_penalty=2, temperature=.5, top_p=.5,
                              diffusion_temperature=.7, cond_free_k=2, diffusion_iterations=200)
 
         down = torchaudio.functional.resample(sample, 24000, 22050)
-        fout_path = os.path.join(outpath, 'old', os.path.basename(line[1]))
-        torchaudio.save(fout_path, down.squeeze(0), 22050)
-
-        down = torchaudio.functional.resample(sample2, 24000, 22050)
-        fout_path = os.path.join(outpath, 'new', os.path.basename(line[1]))
+        fout_path = os.path.join(outpath, os.path.basename(line[1]))
         torchaudio.save(fout_path, down.squeeze(0), 22050)
 
         recorder.write(f'{transcript}\t{fout_path}\n')

read.py

@@ -48,9 +48,10 @@ if __name__ == '__main__':
     parser = argparse.ArgumentParser()
     parser.add_argument('-textfile', type=str, help='A file containing the text to read.', default="data/riding_hood.txt")
     parser.add_argument('-voice', type=str, help='Use a preset conditioning voice (defined above). Overrides cond_path.', default='dotrice')
-    parser.add_argument('-num_samples', type=int, help='How many total outputs the autoregressive transformer should produce.', default=256)
+    parser.add_argument('-num_samples', type=int, help='How many total outputs the autoregressive transformer should produce.', default=512)
     parser.add_argument('-batch_size', type=int, help='How many samples to process at once in the autoregressive model.', default=16)
     parser.add_argument('-output_path', type=str, help='Where to store outputs.', default='results/longform/')
+    parser.add_argument('-generation_preset', type=str, help='Preset to use for generation', default='intelligible')
     args = parser.parse_args()
     os.makedirs(args.output_path, exist_ok=True)
 
@@ -67,7 +68,7 @@ if __name__ == '__main__':
         for cond_path in cond_paths:
             c = load_audio(cond_path, 22050)
             conds.append(c)
-        gen = tts.tts(text, conds, num_autoregressive_samples=args.num_samples, temperature=.7, top_p=.7)
+        gen = tts.tts_with_preset(text, conds, preset=args.generation_preset, num_autoregressive_samples=args.num_samples)
         torchaudio.save(os.path.join(args.output_path, f'{j}.wav'), gen.squeeze(0).cpu(), 24000)
 
         priors.append(torchaudio.functional.resample(gen, 24000, 22050).squeeze(0))