modified how conditional latents are computed (before, it just happened to only bother reading the first 102400/24000=4.26 seconds per audio input, now it will chunk it all to compute latents)

README.md

@@ -115,6 +115,8 @@ To save you from headaches, I strongly recommend playing around with shorter sen
 
 As a quick optimization, I modified the script to where the `conditional_latents` are saved after loading voice samples, and subsequent uses will load that file directly (at the cost of not returning the `Sample voice` to the web UI). If there's voice samples that have a modification time newer than this cached file, it'll skip loading it and load the normal WAVs instead.
 
+**!**NOTE**!**: cached `latents.pth` files generated before 2023.02.05 will be ignored, due to a change in computing the conditiona latents. This *should* help bump up voice cloning quality. Apologies for the inconvenience.
+
 ## Example(s)
 
 Below are some outputs I deem substantial enough to share. As I continue delving into TorToiSe, I'll supply more examples and the values I use.

app.py

@@ -10,7 +10,9 @@ from tortoise.api import TextToSpeech
 from tortoise.utils.audio import load_audio, load_voice, load_voices
 from tortoise.utils.text import split_and_recombine_text
 
-def generate(text, delimiter, emotion, prompt, voice, mic_audio, preset, seed, candidates, num_autoregressive_samples, diffusion_iterations, temperature, diffusion_sampler, breathing_room, progress=gr.Progress()):
+def generate(text, delimiter, emotion, prompt, voice, mic_audio, preset, seed, candidates, num_autoregressive_samples, diffusion_iterations, temperature, diffusion_sampler, breathing_room, experimentals, progress=gr.Progress()):
+    print(experimentals)
+
     if voice != "microphone":
         voices = [voice]
     else:
@@ -27,7 +29,7 @@ def generate(text, delimiter, emotion, prompt, voice, mic_audio, preset, seed, c
     
     if voice_samples is not None:
         sample_voice = voice_samples[0]
-        conditioning_latents = tts.get_conditioning_latents(voice_samples)
+        conditioning_latents = tts.get_conditioning_latents(voice_samples, progress=progress)
         torch.save(conditioning_latents, os.path.join(f'./tortoise/voices/{voice}/', f'latents.pth'))
         voice_samples = None
     else:
@@ -54,6 +56,8 @@ def generate(text, delimiter, emotion, prompt, voice, mic_audio, preset, seed, c
         'diffusion_sampler': diffusion_sampler,
         'breathing_room': breathing_room,
         'progress': progress,
+        'half_p': "Half Precision" in experimentals,
+        'cond_free': "Conditioning-Free" in experimentals,
     }
 
     if delimiter == "\\n":
@@ -216,6 +220,8 @@ def main():
                     type="value",
                 )
 
+                experimentals = gr.CheckboxGroup(["Half Precision", "Conditioning-Free"], value=[False, True], label="Experimental Flags")
+
                 preset.change(fn=update_presets,
                     inputs=preset,
                     outputs=[
@@ -246,7 +252,8 @@ def main():
                         diffusion_iterations,
                         temperature,
                         diffusion_sampler,
-                        breathing_room
+                        breathing_room,
+                        experimentals,
                     ],
                     outputs=[selected_voice, output_audio, usedSeed],
                 )

start.bat

@@ -1,3 +1,4 @@
 call .\tortoise-venv\Scripts\activate.bat
-py .\app.py
+python .\app.py
 deactivate
+pause

tortoise/api.py

@@ -284,7 +284,7 @@ class TextToSpeech:
         if self.minor_optimizations:
             self.cvvp = self.cvvp.to(self.device)
 
-    def get_conditioning_latents(self, voice_samples, return_mels=False):
+    def get_conditioning_latents(self, voice_samples, return_mels=False, verbose=False, progress=None, enforced_length=102400):
         """
         Transforms one or more voice_samples into a tuple (autoregressive_conditioning_latent, diffusion_conditioning_latent).
         These are expressive learned latents that encode aspects of the provided clips like voice, intonation, and acoustic
@@ -303,14 +303,18 @@ class TextToSpeech:
             auto_conds = torch.stack(auto_conds, dim=1)
             
             diffusion_conds = []
-            for sample in voice_samples:
+
+            for sample in tqdm_override(voice_samples, verbose=verbose, progress=progress, desc="Computing conditioning latents..."):
                 # The diffuser operates at a sample rate of 24000 (except for the latent inputs)
                 sample = torchaudio.functional.resample(sample, 22050, 24000)
-                sample = pad_or_truncate(sample, 102400)
+                chunks = torch.chunk(sample, int(sample.shape[-1] / enforced_length) + 1, dim=1)
-                cond_mel = wav_to_univnet_mel(sample.to(self.device), do_normalization=False, device=self.device)
-                diffusion_conds.append(cond_mel)
-            diffusion_conds = torch.stack(diffusion_conds, dim=1)
                 
+                for chunk in chunks:
+                    chunk = pad_or_truncate(chunk, enforced_length)
+                    cond_mel = wav_to_univnet_mel(chunk.to(self.device), do_normalization=False, device=self.device)
+                    diffusion_conds.append(cond_mel)
+
+            diffusion_conds = torch.stack(diffusion_conds, dim=1)
 
             if self.minor_optimizations:
                 auto_latent = self.autoregressive.get_conditioning(auto_conds)
@@ -372,6 +376,7 @@ class TextToSpeech:
             diffusion_iterations=100, cond_free=True, cond_free_k=2, diffusion_temperature=1.0,
             diffusion_sampler="P",
             breathing_room=8,
+            half_p=False,
             progress=None,
             **hf_generate_kwargs):
         """
@@ -446,6 +451,7 @@ class TextToSpeech:
             if not self.minor_optimizations:
                 self.autoregressive = self.autoregressive.to(self.device)
             
+            with torch.autocast(device_type='cuda', dtype=torch.float16, enabled=half_p):
                 for b in tqdm_override(range(num_batches), verbose=verbose, progress=progress, desc="Generating autoregressive samples"):
                     codes = self.autoregressive.inference_speech(auto_conditioning, text_tokens,
                                                                  do_sample=True,
@@ -462,6 +468,7 @@ class TextToSpeech:
 
             clip_results = []
 
+            with torch.autocast(device_type='cuda', dtype=torch.float16, enabled=half_p):
                 if not self.minor_optimizations:
                     self.autoregressive = self.autoregressive.cpu()
                     self.clvp = self.clvp.to(self.device)

tortoise/utils/audio.py

@@ -108,8 +108,11 @@ def load_voice(voice, extra_voice_dirs=[], load_latents=True):
     voices = []
     latent = None
     for file in paths:
-        if file[-4:] == ".pth":
+        if file == "cond_latents.pth":
             latent = file
+        elif file[-4:] == ".pth":
+            {}
+            # noop
         else:
             voices.append(file)
             mtime = max(mtime, os.path.getmtime(file))

update.bat

@@ -0,0 +1,7 @@
+git pull
+python -m venv tortoise-venv
+call .\tortoise-venv\Scripts\activate.bat
+python -m pip install --upgrade pip
+python -m pip install -r ./requirements.txt
+deactivate
+pause