[Feature] Add Faster Whisper (#365)

* Add quick start ipynb

* Remove redundant output

* Fix docs

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* [Feature] Add Fast Whisper

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Different suffix

* Different audio format

* Fix README.md for ja docs

* Fix ZH docs

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

.gitignore

@@ -26,3 +26,4 @@ asr-label*
 /demo-audios
 ref_data*
 /example
+/faster_whisper

README.md

@@ -17,35 +17,43 @@
 </a>
 </div>
 
-This codebase and all models are released under CC-BY-NC-SA-4.0 License. Please refer to [LICENSE](LICENSE) for more details. 
+This codebase and all models are released under CC-BY-NC-SA-4.0 License. Please refer to [LICENSE](LICENSE) for more details.
 
 此代码库及模型根据 CC-BY-NC-SA-4.0 许可证发布。请参阅 [LICENSE](LICENSE) 了解更多细节.
 
 ## Disclaimer / 免责声明
+
 We do not hold any responsibility for any illegal usage of the codebase. Please refer to your local laws about DMCA and other related laws.  
 我们不对代码库的任何非法使用承担任何责任. 请参阅您当地关于 DMCA (数字千年法案) 和其他相关法律法规.
 
 ## Online Demo
-[Fish Audio](https://fish.audio)   
+
+[Fish Audio](https://fish.audio)
 
 ## Quick Start
+
 [inference.ipynb](https://nbviewer.org/github/AnyaCoder/fish-speech/blob/main/inference.ipynb)
 
 ## Videos
+
 #### Demo Video: https://www.bilibili.com/video/BV1wz421B71D
+
 #### Tech slides Video: https://www.bilibili.com/video/BV1zJ4m1K7cj
 
 ## Documents / 文档
+
 - [English](https://speech.fish.audio/en/)
 - [中文](https://speech.fish.audio/)
-- [日本語](https://speech.fish.audio/)
+- [日本語](https://speech.fish.audio/ja/)
 
 ## Samples / 例子
+
 - [English](https://speech.fish.audio/en/samples/)
 - [中文](https://speech.fish.audio/samples/)
 - [日本語](https://speech.fish.audio/ja/samples/)
 
 ## Credits / 鸣谢
+
 - [VITS2 (daniilrobnikov)](https://github.com/daniilrobnikov/vits2)
 - [Bert-VITS2](https://github.com/fishaudio/Bert-VITS2)
 - [GPT VITS](https://github.com/innnky/gpt-vits)

docs/zh/index.md

@@ -67,7 +67,7 @@ Windows 非专业用户可考虑以下为免 Linux 环境的基础运行方法
                   </p>
                </ul>
             </li>
-            <li>インストール <a href="https://developer.nvidia.com/cuda-12-1-0-download-archive?target_os=Windows&target_arch=x86_64">CUDA Toolkit 12</a></li>
+            <li>安装 <a href="https://developer.nvidia.com/cuda-12-1-0-download-archive?target_os=Windows&target_arch=x86_64">CUDA Toolkit 12</a></li>
       </ol>
    </li>
    <li>双击 start.bat, 进入 Fish-Speech 训练推理配置 WebUI 页面。

fish_speech/webui/manage.py

@@ -727,12 +727,10 @@ def llama_quantify(llama_weight, quantify_mode):
                     )
                     label_model = gr.Dropdown(
                         label=i18n("Whisper Model"),
-                        info=i18n(
-                            "Use large for 10G+ GPU, medium for 5G, small for 2G"
-                        ),
-                        choices=["large", "medium", "small"],
-                        value="small",
-                        interactive=True,
+                        info=i18n("Faster Whisper, Up to 5g GPU memory usage"),
+                        choices=["large-v3"],
+                        value="large-v3",
+                        interactive=False,
                     )
                     label_radio = gr.Dropdown(
                         label=i18n("Optional Label Language"),
@@ -744,6 +742,7 @@ def llama_quantify(llama_weight, quantify_mode):
                             (i18n("English"), "EN"),
                             (i18n("Japanese"), "JA"),
                             (i18n("Disabled"), "IGNORE"),
+                            (i18n("auto"), "auto"),
                         ],
                         value="IGNORE",
                         interactive=True,

pyproject.toml

@@ -36,7 +36,9 @@ dependencies = [
     "samplerate>=0.2.1",
     "resampy>=0.4.3",
     "einx[torch]==0.2.2",
-    "zstandard>=0.22.0"
+    "zstandard>=0.22.0",
+    "pydub",
+    "faster_whisper",
 ]
 
 [project.optional-dependencies]

start.bat

@@ -5,7 +5,7 @@ set USE_MIRROR=true
 set PYTHONPATH=%~dp0
 set PYTHON_CMD=%cd%\fishenv\env\python
 set API_FLAG_PATH=%~dp0API_FLAGS.txt
-
+set KMP_DUPLICATE_LIB_OK=TRUE
 
 setlocal enabledelayedexpansion
 

tools/whisper_asr.py

@@ -22,93 +22,157 @@
 Note: Be aware of your audio sample rate, which defaults to 44.1kHz.
 """
 
+import re
 from pathlib import Path
 
 import click
-import librosa
 import soundfile as sf
-import whisper
+from faster_whisper import WhisperModel
 from loguru import logger
-from merge_asr_files import merge_and_delete_files
+from pydub import AudioSegment
 from tqdm import tqdm
 
 from fish_speech.utils.file import AUDIO_EXTENSIONS, list_files
 
 
 @click.command()
-@click.option("--model-size", default="large", help="Size of the Whisper model")
+@click.option("--model-size", default="large-v3", help="Size of the Whisper model")
+@click.option(
+    "--compute-type",
+    default="float16",
+    help="Computation Precision of the Whisper model [float16 / int8_float16 / int8]",
+)
 @click.option("--audio-dir", required=True, help="Directory containing audio files")
 @click.option(
     "--save-dir", required=True, help="Directory to save processed audio files"
 )
 @click.option(
     "--sample-rate",
-    default=None,
+    default=44100,
     type=int,
     help="Output sample rate, default to input sample rate",
 )
-@click.option("--device", default="cuda", help="Device to use")
-@click.option("--language", default="ZH", help="Language of the transcription")
-def main(model_size, audio_dir, save_dir, sample_rate, device, language):
-    logger.info("Loading / Downloading OpenAI Whisper model...")
-    model = whisper.load_model(
-        name=model_size,
+@click.option("--device", default="cuda", help="Device to use [cuda / cpu]")
+@click.option("--language", default="auto", help="Language of the transcription")
+def main(model_size, compute_type, audio_dir, save_dir, sample_rate, device, language):
+    logger.info("Loading / Downloading Faster Whisper model...")
+
+    model = WhisperModel(
+        model_size,
         device=device,
-        download_root=str(Path(".cache/whisper").resolve()),
+        compute_type=compute_type,
+        download_root="faster_whisper",
     )
+
     logger.info("Model loaded.")
 
     save_path = Path(save_dir)
     save_path.mkdir(parents=True, exist_ok=True)
-    original_files = []
+
     audio_files = list_files(
         path=audio_dir, extensions=AUDIO_EXTENSIONS, recursive=True
     )
+
+    numbered_suffix_pattern = re.compile(r"-\d{3}$")
+
     for file_path in tqdm(audio_files, desc="Processing audio file"):
         file_stem = file_path.stem
         file_suffix = file_path.suffix
 
         rel_path = Path(file_path).relative_to(audio_dir)
         (save_path / rel_path.parent).mkdir(parents=True, exist_ok=True)
 
-        if (save_path / rel_path.parent / f"{rel_path.stem}.wav").exists() and (
-            save_path / rel_path.parent / f"{rel_path.stem}.lab"
-        ).exists():
+        # Skip files that already have a .lab file or a -{3-digit number} suffix
+        numbered_suffix = numbered_suffix_pattern.search(file_stem)
+        lab_file = file_path.with_suffix(".lab")
+
+        if numbered_suffix and lab_file.exists():
+            continue
+
+        if not numbered_suffix and lab_file.with_stem(lab_file.stem + "-001").exists():
+            if file_path.exists():
+                file_path.unlink()
             continue
 
-        audio, sr = librosa.load(file_path, sr=sample_rate, mono=False)
-        transcription = model.transcribe(str(file_path), language=language)
+        audio = AudioSegment.from_file(file_path)
 
-        for segment in transcription.get("segments", []):
-            id, text, start, end = (
-                segment["id"],
-                segment["text"],
-                segment["start"],
-                segment["end"],
-            )
+        segments, info = model.transcribe(
+            file_path, beam_size=5, language=None if language == "auto" else language
+        )
 
-            extract = audio[..., int(start * sr) : int(end * sr)]
-            audio_save_path = (
-                save_path / rel_path.parent / f"{file_stem}-{id}{file_suffix}"
+        print(
+            "Detected language '%s' with probability %f"
+            % (info.language, info.language_probability)
+        )
+        print("Total len(ms): ", len(audio))
+
+        for segment in segments:
+            id, start, end, text = (
+                segment.id,
+                segment.start,
+                segment.end,
+                segment.text,
             )
-            sf.write(
-                audio_save_path,
-                extract,
-                samplerate=sr,
+            print("Segment %03d [%.2fs -> %.2fs] %s" % (id, start, end, text))
+            start_ms = int(start * 1000)
+            end_ms = int(end * 1000) + 200  # add 0.2s avoid truncating
+            segment_audio = audio[start_ms:end_ms]
+            audio_save_path = (
+                save_path / rel_path.parent / f"{file_stem}-{id:03d}{file_suffix}"
             )
-            original_files.append(audio_save_path)
+            segment_audio.export(audio_save_path, format=file_suffix[1:])
+            print(f"Exported {audio_save_path}")
 
-            transcript_save_path = save_path / rel_path.parent / f"{file_stem}-{id}.lab"
+            transcript_save_path = (
+                save_path / rel_path.parent / f"{file_stem}-{id:03d}.lab"
+            )
             with open(
                 transcript_save_path,
                 "w",
                 encoding="utf-8",
             ) as f:
-                f.write(text)
-            original_files.append(transcript_save_path)
+                f.write(segment.text)
 
-    merge_and_delete_files(save_dir, original_files)
+        file_path.unlink()
 
 
 if __name__ == "__main__":
     main()
+    exit(0)
+
+    audio = AudioSegment.from_wav(
+        r"D:\PythonProject\原神语音中文\胡桃\vo_hutao_draw_appear.wav"
+    )
+
+    model_size = "large-v3"
+
+    model = WhisperModel(
+        model_size,
+        device="cuda",
+        compute_type="float16",
+        download_root="faster_whisper",
+    )
+
+    segments, info = model.transcribe(
+        r"D:\PythonProject\原神语音中文\胡桃\vo_hutao_draw_appear.wav",
+        beam_size=5,
+    )
+
+    print(
+        "Detected language '%s' with probability %f"
+        % (info.language, info.language_probability)
+    )
+    print("Total len(ms): ", len(audio))
+
+    for i, segment in enumerate(segments):
+        print(
+            "Segment %03d [%.2fs -> %.2fs] %s"
+            % (i, segment.start, segment.end, segment.text)
+        )
+        start_ms = int(segment.start * 1000)
+        end_ms = int(segment.end * 1000)
+        segment_audio = audio[start_ms:end_ms]
+        segment_audio.export(f"segment_{i:03d}.wav", format="wav")
+        print(f"Exported segment_{i:03d}.wav")
+
+    print("All segments have been exported.")