change from chaterbox to gwen-tts

pyproject.toml

@@ -30,11 +30,11 @@ dependencies = [
   "pinyin-tone-converter",
   "genanki",
   "argostranslate",
-  "chatterbox-tts",
   "torch",
   "torchaudio",
   "torchcodec",
   "python-dotenv",
+  "qwen-tts",
 
 ]
 

src/anki_hsk_creator/__init__.py

@@ -18,3 +18,6 @@ HF_TOKEN = os.environ.get("HF_TOKEN")
 default_path = Path.home() / "anki-hsk-creator-data"
 DATA_FOLDER = Path(os.environ.get("DATA_FOLDER", default_path))
 DATA_FOLDER.mkdir(exist_ok=True, parents=True)
+default_gwen = Path.home() / "Gwen"
+GWEN_FOLDER = Path(os.environ.get("GWEN_FOLDER", default_gwen))
+GWEN_FOLDER.mkdir(exist_ok=True, parents=True)

src/anki_hsk_creator/constants.py

@@ -4,7 +4,7 @@
 import importlib.resources
 
 # Local
-from . import DATA_FOLDER
+from . import DATA_FOLDER, GWEN_FOLDER
 
 # Resources
 CCCEDICT_PATH = importlib.resources.files("anki_hsk_creator").joinpath("cedict_ts.u8")
@@ -16,16 +16,13 @@ OUTPUT = DATA_FOLDER / "output"
 OUTPUT.mkdir(exist_ok=True, parents=True)
 RESOURCES = DATA_FOLDER / "resources"
 RESOURCES.mkdir(exist_ok=True, parents=True)
+GWEN_TTS = GWEN_FOLDER / "Qwen3-TTS-12Hz-1.7B-CustomVoice"
 
 # File Types
 PHRASES_TYPE = ".phrases"
 DICT_TYPE = ".dictionary"
 DICTATION_TYPE = ".dictation"
 
-# time dalays
-WORD = 0.1
-COMMA = 0.5
-POINT = 1.0
 
 class LANGUAGES:
     """Available laguages for translation"""
@@ -37,7 +34,6 @@ class LANGUAGES:
     RU = "ru"
     TR = "tr"
     TH = "th"
-
     AvailableLanguages = (EN, ES, FR, RU, TR, TH)
     LanguageNames = {
         EN: "English",

src/anki_hsk_creator/proccessor.py

@@ -8,7 +8,7 @@ import argostranslate.translate
 import torchaudio
 
 # Local
-from .constants import LANGUAGES, WORD, COMMA
+from .constants import LANGUAGES
 from .utility import CCCEDICT, TTS, DictionaryResult, ProcessFile, TranslationResult
 
 # Constants
@@ -31,10 +31,8 @@ def dictation_process(
         audio_line = audio_line.replace("，", "，。。。]")
         audio_path = process_file.resources / f"N{n:03n}.wav"
         if not audio_path.exists():
-            audio = TTS.MODEL.generate(
+            wavs, sr = TTS.generate(f"{audio_line}。")
-                f"{audio_line}。", language_id=LANGUAGES.CN, **TTS.DEFAULTS
+            torchaudio.save(audio_path, wavs[0], sr)
-            )
-            torchaudio.save(audio_path, audio, TTS.MODEL.sr)
         translated = argostranslate.translate.translate(
             line, LANGUAGES.CN, process_file.language_id
         )
@@ -56,10 +54,8 @@ def translator_process(
         audio_line = audio_line.replace("，", "，。。。]")
         audio_path = process_file.resources / f"N{n:03n}.wav"
         if not audio_path.exists():
-            audio = TTS.MODEL.generate(
+            wavs, sr = TTS.generate(f"{audio_line}。")
-                f"{audio_line}。", language_id=LANGUAGES.CN, **TTS.DEFAULTS
+            torchaudio.save(audio_path, wavs[0], sr)
-            )
-            torchaudio.save(audio_path, audio, TTS.MODEL.sr)
         translated = argostranslate.translate.translate(
             line, LANGUAGES.CN, process_file.language_id
         )
@@ -121,10 +117,8 @@ def dictionary_process(process_file: ProcessFile) -> list[DictionaryResult]:
         for line in reader:
             audio_path = process_file.resources / f"{line['pinyin']}.wav"
             if not audio_path.exists():
-                audio = TTS.MODEL.generate(
+                wavs, sr = TTS.generate(f"{line['simplified']}。")
-                    f"{line['simplified']}。", language_id=LANGUAGES.CN, **TTS.DEFAULTS
+                torchaudio.save(audio_path, wavs[0], sr)
-                )
-                torchaudio.save(audio_path, audio, TTS.MODEL.sr)
             print(line)
             result = DictionaryResult(
                 **line, audio_path=audio_path, language_id=process_file.language_id

src/anki_hsk_creator/utility.py

@@ -12,10 +12,10 @@ import argostranslate.package
 import argostranslate.translate
 import torch
 from cedict_utils.cedict import CedictEntry, CedictParser
-from chatterbox.mtl_tts import ChatterboxMultilingualTTS
+from qwen_tts import Qwen3TTSModel
 
 # Local
-from .constants import CCCEDICT_PATH, INPUT, LANGUAGES, OUTPUT, RESOURCES
+from .constants import CCCEDICT_PATH, GWEN_TTS, INPUT, LANGUAGES, OUTPUT, RESOURCES
 
 # Static Clases
 
@@ -50,10 +50,12 @@ class TRANS:
                     if out_package.to_code == to_code:
                         if in_package.to_code == out_package.from_code:
                             print(
-                                f"Check in_package {in_package.from_code} {in_package.to_code}"
+                                f"Check in_package {in_package.from_code}"
+                                f"{in_package.to_code}"
                             )
                             print(
-                                f"Check out_package {out_package.from_code} {out_package.to_code}"
+                                f"Check out_package {out_package.from_code}"
+                                f"{out_package.to_code}"
                             )
                             packages_to_install.append(in_package)
                             packages_to_install.append(out_package)
@@ -137,7 +139,11 @@ class TTS:
 
     MODEL = None
     DEVICE = None
-    DEFAULTS = {"cfg_weight": 0.2, "exaggeration": 0.8}
+    DEFAULTS = {
+        "language": "Chinese",
+        "speaker": "Uncle_Fu",
+        "instruct": "语速缓慢而审慎，每个音节的语调都拿捏得恰到好处，宛如教授在指导新生。",
+    }
 
     @staticmethod
     def create_tts():
@@ -151,10 +157,18 @@ class TTS:
             else:
                 TTS.DEVICE = "cpu"
         if TTS.MODEL is None:
-            TTS.MODEL = ChatterboxMultilingualTTS.from_pretrained(
+            TTS.MODEL = Qwen3TTSModel.from_pretrained(
-                device=TTS.DEVICE, t3_model="v3"
+                GWEN_TTS,
+                device_map=TTS.DEVICE,
+                dtype=torch.bfloat16,
+                # attn_implementation="flash_attention_2",
             )
 
+    @staticmethod
+    def generate(text: str):
+        """Generates a Waw using the defaulst values"""
+        return TTS.MODEL.generate_custom_voice(text=text, **TTS.DEFAULTS)
+
 
 # Clases