updating utils

inference_utils.py

@@ -0,0 +1,71 @@
+import re
+from tortoise.api import TextToSpeech
+from tortoise.utils.audio import load_voice
+
+def clean_text(text: str, target_len: int = 200, max_len: int = 300) -> list[str]:
+    # remove double new line, redundant whitespace, convert non-ascii quotes to ascii quotes
+    text = re.sub(r"\n\n+", r"\n", text)
+    text = re.sub(r"\s+", r" ", text)
+    text = re.sub(r"[“”]", '"', text)
+
+    # split text into sentences, keep quotes together
+    sentences = re.split(r'(?<=[.!?])\s+(?=(?:[^"]*"[^"]*")*[^"]*$)', text)
+
+    # recombine sentences into chunks of desired length
+    chunks = []
+    chunk = ""
+    for sentence in sentences:
+        if len(chunk) + len(sentence) > target_len:
+            chunks.append(chunk)
+            chunk = ""
+        chunk += sentence + " "
+        if len(chunk) > max_len:
+            chunks.append(chunk)
+            chunk = ""
+    if chunk:
+        chunks.append(chunk)
+
+    # clean up chunks, remove leading/trailing whitespace, remove empty/unless chunks
+    chunks = [s.strip() for s in chunks]
+    chunks = [s for s in chunks if s and not re.match(r"^[\s\.,;:!?]*$", s)]
+
+    return chunks
+
+
+def process_textfile(file_path: str) -> list[str]:
+    with open(file_path, "r", encoding="utf-8") as f:
+        text = " ".join([l for l in f.readlines()])
+    text = clean_text(text)
+    return text
+
+def tts(paper_name: str):
+    # load tts model
+    tts = TextToSpeech(
+        autoregressive_model_path="./ai-voice-cloning/training/GlaDOS/finetune/models/5304_gpt.pth"
+    )
+    voice = "GlaDOS"
+    voice_samples, conditioning_latents = load_voice(
+        voice, extra_voice_dirs="./ai-voice-cloning/voices"
+    )
+
+    # process text file
+    texts = process_textfile(f"./llm/scripts/{paper_name}.txt")
+
+    # generate audio for each chunk of text
+    all_audio_chunks = []
+    for i, text in enumerate(texts):
+        gen = tts.tts(
+            text=text,
+            voice=voice,
+            voice_samples=voice_samples,
+            conditioning_latents=conditioning_latents,
+        )
+        torchaudio.save(f"./audio/raw/{i}.wav", gen.squeeze(0).cpu(), 24000)
+
+        all_audio_chunks.append(gen)
+
+    # concatenate all audio chunks
+    full_audio = torch.cat(all_audio_chunks, dim=-1)
+    torchaudio.save(f"./audio/raw/{paper_name}.wav", full_audio, 24000)
+
+print("here")

setup-cuda.sh

@@ -2,10 +2,6 @@
 # get local dependencies
 git submodule init
 git submodule update --remote
-# setup venv
-python3 -m venv venv
-source ./venv/bin/activate
-python3 -m pip install --upgrade pip # just to be safe
 # CUDA
 pip3 install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu118
 # install requirements
@@ -16,5 +12,3 @@ python3 -m pip install -e ./modules/dlas/ # install DLAS
 python3 -m pip install -r ./requirements.txt # install local requirements
 
 rm *.bat
-
-deactivate

tortoise_utils.py

@@ -0,0 +1,73 @@
+import re
+from tortoise.api import TextToSpeech
+from tortoise.utils.audio import load_voice
+
+def clean_text(text: str, target_len: int = 200, max_len: int = 300) -> list[str]:
+    # remove double new line, redundant whitespace, convert non-ascii quotes to ascii quotes
+    text = re.sub(r"\n\n+", r"\n", text)
+    text = re.sub(r"\s+", r" ", text)
+    text = re.sub(r"[“”]", '"', text)
+
+    # split text into sentences, keep quotes together
+    sentences = re.split(r'(?<=[.!?])\s+(?=(?:[^"]*"[^"]*")*[^"]*$)', text)
+
+    # recombine sentences into chunks of desired length
+    chunks = []
+    chunk = ""
+    for sentence in sentences:
+        if len(chunk) + len(sentence) > target_len:
+            chunks.append(chunk)
+            chunk = ""
+        chunk += sentence + " "
+        if len(chunk) > max_len:
+            chunks.append(chunk)
+            chunk = ""
+    if chunk:
+        chunks.append(chunk)
+
+    # clean up chunks, remove leading/trailing whitespace, remove empty/unless chunks
+    chunks = [s.strip() for s in chunks]
+    chunks = [s for s in chunks if s and not re.match(r"^[\s\.,;:!?]*$", s)]
+
+    return chunks
+
+
+def process_textfile(file_path: str) -> list[str]:
+    with open(file_path, "r", encoding="utf-8") as f:
+        text = " ".join([l for l in f.readlines()])
+    text = clean_text(text)
+    return text
+
+def tts(file_path: str):
+    # load tts model
+    # ADD PATH
+    tts = TextToSpeech(
+        autoregressive_model_path="./ai-voice-cloning/training/"
+    )
+    voice = "Lex"
+    voice_samples, conditioning_latents = load_voice(
+        voice, extra_voice_dirs="./ai-voice-cloning/voices"
+    )
+
+    # process text file
+    texts = process_textfile(file_path)
+
+    # generate audio for each chunk of text
+    all_audio_chunks = []
+    for i, text in enumerate(texts):
+        gen = tts.tts(
+            text=text,
+            voice=voice,
+            voice_samples=voice_samples,
+            conditioning_latents=conditioning_latents,
+        )
+        torchaudio.save(f"./audio/raw/{i}.wav", gen.squeeze(0).cpu(), 24000)
+
+        all_audio_chunks.append(gen)
+
+    book_name_ext = os.path.basename(file_path)
+    paper_name = os.path.splitext(book_name_ext)[0]
+
+    # concatenate all audio chunks
+    full_audio = torch.cat(all_audio_chunks, dim=-1)
+    torchaudio.save(f"./audio/raw/{paper_name}.wav", full_audio, 24000)