New version using faster-whisper with more model parameters exposed

.gitignore

@@ -1,5 +1,5 @@
 dist
 .DS_Store
 *.egg-info
-auto_subtitle/__pycache__
 build
+__pycache__

LICENSE

@@ -1,6 +1,7 @@
 MIT License
 
-Copyright (c) 2022 Miguel Piedrafita <soy@miguelpiedrafita.com>
+Copyright (c) 2022-2024 Miguel Piedrafita <soy@miguelpiedrafita.com>
+Copyright (c) 2024 Sergey Chernyaev <schernyae@gmail.com>
 
 Permission is hereby granted, free of charge, to any person obtaining a copy
 of this software and associated documentation files (the "Software"), to deal

README.md

@@ -1,12 +1,12 @@
 # Automatic subtitles in your videos
 
-This repository uses `ffmpeg` and [OpenAI's Whisper](https://openai.com/blog/whisper) to automatically generate and overlay subtitles on any video.
+This repository uses `ffmpeg` and [OpenAI's Whisper](https://openai.com/blog/whisper) ([faster-whisper](https://github.com/SYSTRAN/faster-whisper) implementation) to automatically generate and overlay subtitles on any video.
 
 ## Installation
 
 To get started, you'll need Python 3.7 or newer. Install the binary by running the following command:
 
-    pip install git+https://github.com/m1guelpf/auto-subtitle.git
+    pip install git+https://github.com/Sirozha1337/faster-auto-subtitle.git@dev
 
 You'll also need to install [`ffmpeg`](https://ffmpeg.org/), which is available from most package managers:
 
@@ -25,19 +25,32 @@ choco install ffmpeg
 
 The following command will generate a `subtitled/video.mp4` file contained the input video with overlayed subtitles.
 
-    auto_subtitle /path/to/video.mp4 -o subtitled/
+    faster_auto_subtitle /path/to/video.mp4 -o subtitled/
 
-The default setting (which selects the `small` model) works well for transcribing English. You can optionally use a bigger model for better results (especially with other languages). The available models are `tiny`, `tiny.en`, `base`, `base.en`, `small`, `small.en`, `medium`, `medium.en`, `large`.
+The default setting (which selects the `small` model) works well for transcribing English. You can optionally use a bigger model for better results (especially with other languages). The available models are `tiny`, `tiny.en`, `base`, `base.en`, `small`, `small.en`, `medium`, `medium.en`, `large`, `large-v1`, `large-v2`, `large-v3`.
-
+    faster_auto_subtitle /path/to/video.mp4 --model medium
-    auto_subtitle /path/to/video.mp4 --model medium
 
 Adding `--task translate` will translate the subtitles into English:
 
-    auto_subtitle /path/to/video.mp4 --task translate
+    faster_auto_subtitle /path/to/video.mp4 --task translate
 
 Run the following to view all available options:
 
-    auto_subtitle --help
+    faster_auto_subtitle --help
+
+## Tips
+
+The tool also exposes a couple of model parameters, that you can tweak to increase accuracy.
+
+Higher `beam_size` usually leads to greater accuraccy, but slows down the process.
+
+Setting higher `no_speech_threshold` could be useful for videos with a lot of background noise to stop Whisper from "hallucinating" subtitles for it.
+
+In my experience settings option `condition_on_previous_text` to `False` dramatically increases accurracy for videos like TV Shows with an intro song at the start. 
+
+You can use `sample_interval` parameter to generate subtitles for a portion of the video to play around with those parameters:
+
+    faster_auto_subtitle /path/to/video.mp4 --model medium --sample_interval 00:05:30-00:07:00 --condition_on_previous_text False --beam_size 6 --no_speech_threshold 0.7
 
 ## License
 

auto_subtitle/cli.py

@@ -1,114 +1,39 @@
-import os
-import ffmpeg
-import whisper
 import argparse
-import warnings
+from faster_whisper import available_models
-import tempfile
+from .main import process
-from .utils import filename, str2bool, write_srt
+from .utils.convert import str2bool, str2timeinterval
-
 
 def main():
     parser = argparse.ArgumentParser(
         formatter_class=argparse.ArgumentDefaultsHelpFormatter)
     parser.add_argument("video", nargs="+", type=str,
                         help="paths to video files to transcribe")
+    parser.add_argument("--audio_channel", default="0",
+                        type=int, help="audio channel index to use")
+    parser.add_argument("--sample_interval", type=str2timeinterval, default=None,
+                        help="generate subtitles for a specific fragment of the video (e.g. 01:02:05-01:03:45)")
     parser.add_argument("--model", default="small",
-                        choices=whisper.available_models(), help="name of the Whisper model to use")
+                        choices=available_models(), help="name of the Whisper model to use")
     parser.add_argument("--output_dir", "-o", type=str,
                         default=".", help="directory to save the outputs")
     parser.add_argument("--output_srt", type=str2bool, default=False,
                         help="whether to output the .srt file along with the video files")
     parser.add_argument("--srt_only", type=str2bool, default=False,
                         help="only generate the .srt file and not create overlayed video")
-    parser.add_argument("--verbose", type=str2bool, default=False,
+    parser.add_argument("--beam_size", type=int, default=5,
-                        help="whether to print out the progress and debug messages")
+                        help="model parameter, tweak to increase accuracy")
-
+    parser.add_argument("--no_speech_threshold", type=float, default=0.6,
+                        help="model parameter, tweak to increase accuracy")
+    parser.add_argument("--condition_on_previous_text", type=str2bool, default=True,
+                        help="model parameter, tweak to increase accuracy")
     parser.add_argument("--task", type=str, default="transcribe", choices=[
                         "transcribe", "translate"], help="whether to perform X->X speech recognition ('transcribe') or X->English translation ('translate')")
     parser.add_argument("--language", type=str, default="auto", choices=["auto","af","am","ar","as","az","ba","be","bg","bn","bo","br","bs","ca","cs","cy","da","de","el","en","es","et","eu","fa","fi","fo","fr","gl","gu","ha","haw","he","hi","hr","ht","hu","hy","id","is","it","ja","jw","ka","kk","km","kn","ko","la","lb","ln","lo","lt","lv","mg","mi","mk","ml","mn","mr","ms","mt","my","ne","nl","nn","no","oc","pa","pl","ps","pt","ro","ru","sa","sd","si","sk","sl","sn","so","sq","sr","su","sv","sw","ta","te","tg","th","tk","tl","tr","tt","uk","ur","uz","vi","yi","yo","zh"], 
     help="What is the origin language of the video? If unset, it is detected automatically.")
 
     args = parser.parse_args().__dict__
-    model_name: str = args.pop("model")
-    output_dir: str = args.pop("output_dir")
-    output_srt: bool = args.pop("output_srt")
-    srt_only: bool = args.pop("srt_only")
-    language: str = args.pop("language")
 
-    os.makedirs(output_dir, exist_ok=True)
+    process(args)
-
-    if model_name.endswith(".en"):
-        warnings.warn(
-            f"{model_name} is an English-only model, forcing English detection.")
-        args["language"] = "en"
-    # if translate task used and language argument is set, then use it
-    elif language != "auto":
-        args["language"] = language
-        
-    model = whisper.load_model(model_name)
-    audios = get_audio(args.pop("video"))
-    subtitles = get_subtitles(
-        audios, output_srt or srt_only, output_dir, lambda audio_path: model.transcribe(audio_path, **args)
-    )
-
-    if srt_only:
-        return
-
-    for path, srt_path in subtitles.items():
-        out_path = os.path.join(output_dir, f"{filename(path)}.mp4")
-
-        print(f"Adding subtitles to {filename(path)}...")
-
-        video = ffmpeg.input(path)
-        audio = video.audio
-
-        ffmpeg.concat(
-            video.filter('subtitles', srt_path, force_style="OutlineColour=&H40000000,BorderStyle=3"), audio, v=1, a=1
-        ).output(out_path).run(quiet=True, overwrite_output=True)
-
-        print(f"Saved subtitled video to {os.path.abspath(out_path)}.")
-
-
-def get_audio(paths):
-    temp_dir = tempfile.gettempdir()
-
-    audio_paths = {}
-
-    for path in paths:
-        print(f"Extracting audio from {filename(path)}...")
-        output_path = os.path.join(temp_dir, f"{filename(path)}.wav")
-
-        ffmpeg.input(path).output(
-            output_path,
-            acodec="pcm_s16le", ac=1, ar="16k"
-        ).run(quiet=True, overwrite_output=True)
-
-        audio_paths[path] = output_path
-
-    return audio_paths
-
-
-def get_subtitles(audio_paths: list, output_srt: bool, output_dir: str, transcribe: callable):
-    subtitles_path = {}
-
-    for path, audio_path in audio_paths.items():
-        srt_path = output_dir if output_srt else tempfile.gettempdir()
-        srt_path = os.path.join(srt_path, f"{filename(path)}.srt")
-        
-        print(
-            f"Generating subtitles for {filename(path)}... This might take a while."
-        )
-
-        warnings.filterwarnings("ignore")
-        result = transcribe(audio_path)
-        warnings.filterwarnings("default")
-
-        with open(srt_path, "w", encoding="utf-8") as srt:
-            write_srt(result["segments"], file=srt)
-
-        subtitles_path[path] = srt_path
-
-    return subtitles_path
 
 
 if __name__ == '__main__':

auto_subtitle/main.py

@@ -0,0 +1,55 @@
+import os
+import warnings
+import tempfile
+from .utils.files import filename, write_srt
+from .utils.ffmpeg import get_audio, overlay_subtitles
+from .utils.whisper import WhisperAI
+
+def process(args: dict):
+    model_name: str = args.pop("model")
+    output_dir: str = args.pop("output_dir")
+    output_srt: bool = args.pop("output_srt")
+    srt_only: bool = args.pop("srt_only")
+    language: str = args.pop("language")
+    sample_interval: str = args.pop("sample_interval")
+    
+    os.makedirs(output_dir, exist_ok=True)
+
+    if model_name.endswith(".en"):
+        warnings.warn(
+            f"{model_name} is an English-only model, forcing English detection.")
+        args["language"] = "en"
+    # if translate task used and language argument is set, then use it
+    elif language != "auto":
+        args["language"] = language
+
+    audios = get_audio(args.pop("video"), args.pop('audio_channel'), sample_interval)
+    subtitles = get_subtitles(
+        audios, output_srt or srt_only, output_dir, model_name, args
+    )
+
+    if srt_only:
+        return
+
+    overlay_subtitles(subtitles, output_dir, sample_interval)
+
+def get_subtitles(audio_paths: list, output_srt: bool, output_dir: str, model_name: str, model_args: dict):
+    model = WhisperAI(model_name, model_args)
+
+    subtitles_path = {}
+
+    for path, audio_path in audio_paths.items():
+        print(
+            f"Generating subtitles for {filename(path)}... This might take a while."
+        )
+        srt_path = output_dir if output_srt else tempfile.gettempdir()
+        srt_path = os.path.join(srt_path, f"{filename(path)}.srt")
+        
+        segments = model.transcribe(audio_path)
+
+        with open(srt_path, "w", encoding="utf-8") as srt:
+            write_srt(segments, file=srt)
+
+        subtitles_path[path] = srt_path
+
+    return subtitles_path

auto_subtitle/utils.py

@@ -1,46 +0,0 @@
-import os
-from typing import Iterator, TextIO
-
-
-def str2bool(string):
-    string = string.lower()
-    str2val = {"true": True, "false": False}
-
-    if string in str2val:
-        return str2val[string]
-    else:
-        raise ValueError(
-            f"Expected one of {set(str2val.keys())}, got {string}")
-
-
-def format_timestamp(seconds: float, always_include_hours: bool = False):
-    assert seconds >= 0, "non-negative timestamp expected"
-    milliseconds = round(seconds * 1000.0)
-
-    hours = milliseconds // 3_600_000
-    milliseconds -= hours * 3_600_000
-
-    minutes = milliseconds // 60_000
-    milliseconds -= minutes * 60_000
-
-    seconds = milliseconds // 1_000
-    milliseconds -= seconds * 1_000
-
-    hours_marker = f"{hours:02d}:" if always_include_hours or hours > 0 else ""
-    return f"{hours_marker}{minutes:02d}:{seconds:02d},{milliseconds:03d}"
-
-
-def write_srt(transcript: Iterator[dict], file: TextIO):
-    for i, segment in enumerate(transcript, start=1):
-        print(
-            f"{i}\n"
-            f"{format_timestamp(segment['start'], always_include_hours=True)} --> "
-            f"{format_timestamp(segment['end'], always_include_hours=True)}\n"
-            f"{segment['text'].strip().replace('-->', '->')}\n",
-            file=file,
-            flush=True,
-        )
-
-
-def filename(path):
-    return os.path.splitext(os.path.basename(path))[0]

auto_subtitle/utils/convert.py

@@ -0,0 +1,82 @@
+from datetime import datetime, timedelta
+
+def str2bool(string):
+    string = string.lower()
+    str2val = {"true": True, "false": False}
+
+    if string in str2val:
+        return str2val[string]
+    else:
+        raise ValueError(
+            f"Expected one of {set(str2val.keys())}, got {string}")
+
+def str2timeinterval(string):
+    if string is None:
+        return None
+    
+    if '-' not in string:
+        raise ValueError(
+            f"Expected time interval HH:mm:ss-HH:mm:ss or HH:mm-HH:mm or ss-ss, got {string}")
+    
+    intervals = string.split('-')
+    if len(intervals) != 2:
+        raise ValueError(
+            f"Expected time interval HH:mm:ss-HH:mm:ss or HH:mm-HH:mm or ss-ss, got {string}")
+
+    start = try_parse_timestamp(intervals[0])
+    end = try_parse_timestamp(intervals[1])
+    if start >= end:
+        raise ValueError(
+            f"Expected time interval end to be higher than start, got {start} >= {end}")
+    
+    return [start, end]
+
+def time_to_timestamp(string):
+    split_time = string.split(':')
+    if len(split_time) == 0 or len(split_time) > 3 or not all([ x.isdigit() for x in split_time ]):
+        raise ValueError(
+            f"Expected HH:mm:ss or HH:mm or ss, got {string}")
+    
+    if len(split_time) == 1:
+        return int(split_time[0])
+    
+    if len(split_time) == 2:
+        return int(split_time[0]) * 60 * 60 + int(split_time[1]) * 60
+    
+    return int(split_time[0]) * 60 * 60 + int(split_time[1]) * 60 + int(split_time[2])
+
+def try_parse_timestamp(string):
+    timestamp = parse_timestamp(string, '%H:%M:%S')
+    if timestamp is not None:
+        return timestamp
+    
+    timestamp = parse_timestamp(string, '%H:%M')
+    if timestamp is not None:
+        return timestamp
+    
+    return parse_timestamp(string, '%S')
+
+def parse_timestamp(string, pattern):
+    try:
+        date = datetime.strptime(string, pattern)
+        delta = timedelta(hours=date.hour, minutes=date.minute, seconds=date.second)
+        return int(delta.total_seconds())
+    except:
+        return None
+
+def format_timestamp(seconds: float, always_include_hours: bool = False):
+    assert seconds >= 0, "non-negative timestamp expected"
+    milliseconds = round(seconds * 1000.0)
+
+    hours = milliseconds // 3_600_000
+    milliseconds -= hours * 3_600_000
+
+    minutes = milliseconds // 60_000
+    milliseconds -= minutes * 60_000
+
+    seconds = milliseconds // 1_000
+    milliseconds -= seconds * 1_000
+
+    hours_marker = f"{hours:02d}:" if always_include_hours or hours > 0 else ""
+    return f"{hours_marker}{minutes:02d}:{seconds:02d},{milliseconds:03d}"
+

auto_subtitle/utils/ffmpeg.py

@@ -0,0 +1,65 @@
+import os
+import ffmpeg
+import tempfile
+from .mytempfile import MyTempFile
+from .files import filename
+
+def get_audio(paths: list, audio_channel_index: int, sample_interval: list):
+    temp_dir = tempfile.gettempdir()
+
+    audio_paths = {}
+
+    for path in paths:
+        print(f"Extracting audio from {filename(path)}...")
+        output_path = os.path.join(temp_dir, f"{filename(path)}.wav")
+
+        ffmpeg_input_args = dict()
+        if sample_interval is not None:
+            ffmpeg_input_args['ss'] = str(sample_interval[0])
+
+        ffmpeg_output_args = dict()
+        ffmpeg_output_args['acodec'] = "pcm_s16le"
+        ffmpeg_output_args['ac'] = "1"
+        ffmpeg_output_args['ar'] = "16k"
+        ffmpeg_output_args['map'] = "0:a:" + str(audio_channel_index)
+        if sample_interval is not None:
+            ffmpeg_output_args['t'] = str(sample_interval[1] - sample_interval[0])
+            
+        ffmpeg.input(path, **ffmpeg_input_args).output(
+            output_path,
+            **ffmpeg_output_args
+        ).run(quiet=True, overwrite_output=True)
+
+        audio_paths[path] = output_path
+
+    return audio_paths
+
+def escape_windows_path(path: str):
+    return path.replace("\\", "/").replace(":", ":").replace(" ", "\\ ").replace("(", "\\(").replace(")", "\\)").replace("[", "\\[").replace("]", "\\]").replace("'", "'\\''")
+
+
+def overlay_subtitles(subtitles: dict, output_dir: str, sample_interval: list):
+    for path, srt_path in subtitles.items():
+        out_path = os.path.join(output_dir, f"{filename(path)}.mp4")
+
+        print(f"Adding subtitles to {filename(path)}...")
+
+        ffmpeg_input_args = dict()
+        if sample_interval is not None:
+            ffmpeg_input_args['ss'] = str(sample_interval[0])
+
+        ffmpeg_output_args = dict()
+        if sample_interval is not None:
+            ffmpeg_output_args['t'] = str(sample_interval[1] - sample_interval[0])
+
+        # HACK: On Windows it's impossible to use absolute subtitle file path with ffmpeg, so we use temp copy instead
+        # see: https://github.com/kkroening/ffmpeg-python/issues/745
+        with MyTempFile(srt_path) as srt_temp:
+            video = ffmpeg.input(path, **ffmpeg_input_args)
+            audio = video.audio
+
+            ffmpeg.concat(
+                video.filter('subtitles', srt_temp.tmp_file_path, force_style="OutlineColour=&H40000000,BorderStyle=3"), audio, v=1, a=1
+            ).output(out_path, **ffmpeg_output_args).run(quiet=True, overwrite_output=True)
+
+        print(f"Saved subtitled video to {os.path.abspath(out_path)}.")

auto_subtitle/utils/files.py

@@ -0,0 +1,17 @@
+import os
+from typing import Iterator, TextIO
+from .convert import format_timestamp
+
+def write_srt(transcript: Iterator[dict], file: TextIO):
+    for i, segment in enumerate(transcript, start=1):
+        print(
+            f"{i}\n"
+            f"{format_timestamp(segment.start, always_include_hours=True)} --> "
+            f"{format_timestamp(segment.end, always_include_hours=True)}\n"
+            f"{segment.text.strip().replace('-->', '->')}\n",
+            file=file,
+            flush=True,
+        )
+
+def filename(path):
+    return os.path.splitext(os.path.basename(path))[0]

auto_subtitle/utils/mytempfile.py

@@ -0,0 +1,18 @@
+import tempfile
+import os
+import shutil
+
+class MyTempFile:
+    def __init__(self, file_path):
+        self.file_path = file_path
+
+    def __enter__(self):
+        self.tmp_file = tempfile.NamedTemporaryFile('w', dir='.', delete=False)
+        self.tmp_file_path = os.path.relpath(self.tmp_file.name, '.')
+        shutil.copyfile(self.file_path, self.tmp_file_path)
+        return self
+
+    def __exit__(self, exc_type, exc_value, exc_traceback):
+        self.tmp_file.close()
+        if os.path.isfile(self.tmp_file_path):
+            os.remove(self.tmp_file_path)

auto_subtitle/utils/whisper.py

@@ -0,0 +1,20 @@
+import warnings
+import faster_whisper
+from tqdm import tqdm
+
+class WhisperAI:
+    def __init__(self, model_name, model_args):
+        self.model = faster_whisper.WhisperModel(model_name, device="cuda", compute_type="float16")
+        self.model_args = model_args
+
+    def transcribe(self, audio_path):
+        warnings.filterwarnings("ignore")
+        segments, info = self.model.transcribe(audio_path, **self.model_args)
+        warnings.filterwarnings("default")
+
+        total_duration = round(info.duration, 2)  # Same precision as the Whisper timestamps.
+
+        with tqdm(total=total_duration, unit=" seconds") as pbar:
+            for segment in segments:
+                yield segment
+                pbar.update(segment.end - segment.start)

requirements.txt

@@ -1 +1,3 @@
-openai-whisper
+faster-whisper==0.10.0
+tqdm==4.56.0
+ffmpeg-python==0.2.0

setup.py

@@ -2,16 +2,18 @@ from setuptools import setup, find_packages
 
 setup(
     version="1.0",
-    name="auto_subtitle",
+    name="faster_auto_subtitle",
     packages=find_packages(),
     py_modules=["auto_subtitle"],
-    author="Miguel Piedrafita",
+    author="Sergey Chernyaev",
     install_requires=[
-        'openai-whisper',
+        'faster-whisper',
+        'tqdm',
+        'ffmpeg-python'
     ],
     description="Automatically generate and embed subtitles into your videos",
     entry_points={
-        'console_scripts': ['auto_subtitle=auto_subtitle.cli:main'],
+        'console_scripts': ['faster_auto_subtitle=auto_subtitle.cli:main'],
     },
     include_package_data=True,
 )