1
0
Fork 0
mirror of https://github.com/morpheus65535/bazarr synced 2024-12-29 11:06:02 +00:00
bazarr/custom_libs/subliminal_patch/providers/whisperai.py
JayZed 5749971d67
Improved whisper provider to not throttle when unsupported audio language is encountered. #2474
As we have noted before, bad input data should be no reason to throttle a provider.
In this case, if the input language was not supported by whisper, we were raising a ValueError that was never caught and causing an error in the whisper provider for which it was throttled.
Instead, we are now detecting this case and logging an error message.
However, given that the input language was not one of the 99 currently known to whisper, it's probably a mislabeled audio track. If the user desired output language is English, then we will tell whisper that the input audio is also English and ask it to transcribe it. Whisper does a very good job of transcribing almost anything to English, so it's worth a try.
This should address the throttling in issue #2474.
2024-04-29 22:11:47 -04:00

361 lines
12 KiB
Python

from __future__ import absolute_import
import logging
import time
from datetime import timedelta
from requests import Session
from subliminal_patch.subtitle import Subtitle
from subliminal_patch.providers import Provider
from subliminal import __short_version__
from subliminal.exceptions import ConfigurationError
from subzero.language import Language
from subliminal.video import Episode, Movie
from babelfish.exceptions import LanguageReverseError
import ffmpeg
import functools
# These are all the languages Whisper supports.
# from whisper.tokenizer import LANGUAGES
whisper_languages = {
"en": "english",
"zh": "chinese",
"de": "german",
"es": "spanish",
"ru": "russian",
"ko": "korean",
"fr": "french",
"ja": "japanese",
"pt": "portuguese",
"tr": "turkish",
"pl": "polish",
"ca": "catalan",
"nl": "dutch",
"ar": "arabic",
"sv": "swedish",
"it": "italian",
"id": "indonesian",
"hi": "hindi",
"fi": "finnish",
"vi": "vietnamese",
"he": "hebrew",
"uk": "ukrainian",
"el": "greek",
"ms": "malay",
"cs": "czech",
"ro": "romanian",
"da": "danish",
"hu": "hungarian",
"ta": "tamil",
"no": "norwegian",
"th": "thai",
"ur": "urdu",
"hr": "croatian",
"bg": "bulgarian",
"lt": "lithuanian",
"la": "latin",
"mi": "maori",
"ml": "malayalam",
"cy": "welsh",
"sk": "slovak",
"te": "telugu",
"fa": "persian",
"lv": "latvian",
"bn": "bengali",
"sr": "serbian",
"az": "azerbaijani",
"sl": "slovenian",
"kn": "kannada",
"et": "estonian",
"mk": "macedonian",
"br": "breton",
"eu": "basque",
"is": "icelandic",
"hy": "armenian",
"ne": "nepali",
"mn": "mongolian",
"bs": "bosnian",
"kk": "kazakh",
"sq": "albanian",
"sw": "swahili",
"gl": "galician",
"mr": "marathi",
"pa": "punjabi",
"si": "sinhala",
"km": "khmer",
"sn": "shona",
"yo": "yoruba",
"so": "somali",
"af": "afrikaans",
"oc": "occitan",
"ka": "georgian",
"be": "belarusian",
"tg": "tajik",
"sd": "sindhi",
"gu": "gujarati",
"am": "amharic",
"yi": "yiddish",
"lo": "lao",
"uz": "uzbek",
"fo": "faroese",
"ht": "haitian creole",
"ps": "pashto",
"tk": "turkmen",
"nn": "nynorsk",
"mt": "maltese",
"sa": "sanskrit",
"lb": "luxembourgish",
"my": "myanmar",
"bo": "tibetan",
"tl": "tagalog",
"mg": "malagasy",
"as": "assamese",
"tt": "tatar",
"haw": "hawaiian",
"ln": "lingala",
"ha": "hausa",
"ba": "bashkir",
"jw": "javanese",
"su": "sundanese",
}
logger = logging.getLogger(__name__)
def set_log_level(newLevel="INFO"):
newLevel = newLevel.upper()
# print(f'WhisperAI log level changing from {logging._levelToName[logger.getEffectiveLevel()]} to {newLevel}')
logger.setLevel(getattr(logging, newLevel))
# initialize to default above
set_log_level()
@functools.lru_cache(2)
def encode_audio_stream(path, ffmpeg_path, audio_stream_language=None):
logger.debug("Encoding audio stream to WAV with ffmpeg")
try:
# This launches a subprocess to decode audio while down-mixing and resampling as necessary.
inp = ffmpeg.input(path, threads=0)
if audio_stream_language:
logger.debug(f"Whisper will only use the {audio_stream_language} audio stream for {path}")
inp = inp[f'a:m:language:{audio_stream_language}']
out, _ = inp.output("-", format="s16le", acodec="pcm_s16le", ac=1, ar=16000) \
.run(cmd=[ffmpeg_path, "-nostdin"], capture_stdout=True, capture_stderr=True)
except ffmpeg.Error as e:
logger.warning(f"ffmpeg failed to load audio: {e.stderr.decode()}")
return None
logger.debug(f"Finished encoding audio stream in {path} with no errors")
return out
def whisper_get_language(code, name):
# Whisper uses an inconsistent mix of alpha2 and alpha3 language codes
try:
return Language.fromalpha2(code)
except LanguageReverseError:
return Language.fromname(name)
def whisper_get_language_reverse(alpha3):
# Returns the whisper language code given an alpha3b language
for wl in whisper_languages:
lan = whisper_get_language(wl, whisper_languages[wl])
if lan.alpha3 == alpha3:
return wl
return None
def language_from_alpha3(lang):
name = Language(lang).name
return name
class WhisperAISubtitle(Subtitle):
'''Whisper AI Subtitle.'''
provider_name = 'whisperai'
hash_verifiable = False
def __init__(self, language, video):
super(WhisperAISubtitle, self).__init__(language)
self.video = video
self.task = None
self.audio_language = None
self.force_audio_stream = None
@property
def id(self):
return self.video.original_name
def get_matches(self, video):
matches = set()
if isinstance(video, Episode):
matches.update(["series", "season", "episode"])
elif isinstance(video, Movie):
matches.update(["title"])
return matches
class WhisperAIProvider(Provider):
'''Whisper AI Provider.'''
languages = set()
for lan in whisper_languages:
languages.update({whisper_get_language(lan, whisper_languages[lan])})
video_types = (Episode, Movie)
def __init__(self, endpoint=None, response=None, timeout=None, ffmpeg_path=None, loglevel=None):
set_log_level(loglevel)
if not endpoint:
raise ConfigurationError('Whisper Web Service Endpoint must be provided')
if not response:
raise ConfigurationError('Whisper Web Service Connection/response timeout must be provided')
if not timeout:
raise ConfigurationError('Whisper Web Service Transcription/translation timeout must be provided')
if not ffmpeg_path:
raise ConfigurationError("ffmpeg path must be provided")
self.endpoint = endpoint.rstrip("/")
self.response = int(response)
self.timeout = int(timeout)
self.session = None
self.ffmpeg_path = ffmpeg_path
def initialize(self):
self.session = Session()
self.session.headers['User-Agent'] = 'Subliminal/%s' % __short_version__
def terminate(self):
self.session.close()
@functools.lru_cache(2048)
def detect_language(self, path) -> Language:
out = encode_audio_stream(path, self.ffmpeg_path)
if out == None:
logger.info(f"Whisper cannot detect language of {path} because of missing/bad audio track")
return None
r = self.session.post(f"{self.endpoint}/detect-language",
params={'encode': 'false'},
files={'audio_file': out},
timeout=(self.response, self.timeout))
logger.debug(f"Whisper detected language of {path} as {r.json()['detected_language']}")
return whisper_get_language(r.json()["language_code"], r.json()["detected_language"])
def query(self, language, video):
if language not in self.languages:
return None
sub = WhisperAISubtitle(language, video)
sub.task = "transcribe"
if video.audio_languages and not (list(video.audio_languages)[0] == "und" and len(video.audio_languages) == 1):
if language.alpha3 in video.audio_languages:
sub.audio_language = language.alpha3
if len(list(video.audio_languages)) > 1:
sub.force_audio_stream = language.alpha3
else:
sub.task = "translate"
eligible_languages = list(video.audio_languages)
if len(eligible_languages) > 1:
if "und" in eligible_languages:
eligible_languages.remove("und")
sub.audio_language = eligible_languages[0]
else:
# We must detect the language manually
detected_lang = self.detect_language(video.original_path)
if detected_lang == None:
sub.task = "error"
# tell the user what is wrong
sub.release_info = "bad/missing audio track - cannot transcribe"
return sub
if detected_lang != language:
sub.task = "translate"
sub.audio_language = detected_lang.alpha3
if sub.task == "translate":
if language.alpha3 != "eng":
logger.debug(f"Translation only possible from {language} to English")
return None
# tell the user what we are about to do
sub.release_info = f"{sub.task} {language_from_alpha3(sub.audio_language)} audio -> {language_from_alpha3(language.alpha3)} SRT"
logger.debug(f"Whisper query: ({video.original_path}): {sub.audio_language} -> {language.alpha3} [TASK: {sub.task}]")
return sub
def list_subtitles(self, video, languages):
subtitles = [self.query(l, video) for l in languages]
return [s for s in subtitles if s is not None]
def download_subtitle(self, subtitle: WhisperAISubtitle):
# Invoke Whisper through the API. This may take a long time depending on the file.
# TODO: This loads the entire file into memory, find a good way to stream the file in chunks
out = None
if subtitle.task != "error":
out = encode_audio_stream(subtitle.video.original_path, self.ffmpeg_path, subtitle.force_audio_stream)
if out == None:
logger.info(f"Whisper cannot process {subtitle.video.original_path} because of missing/bad audio track")
subtitle.content = None
return
logger.debug(f'Audio stream length (in WAV format) is {len(out):,} bytes')
if subtitle.task == "transcribe":
output_language = subtitle.audio_language
else:
output_language = "eng"
input_language = whisper_get_language_reverse(subtitle.audio_language)
if input_language is None:
if output_language == "eng":
# guess that audio track is mislabelled English and let whisper try to transcribe it
input_language = "en"
subtitle.task = "transcribe"
logger.info(f"Whisper treating unsupported audio track language: '{subtitle.audio_language}' as English")
else:
logger.info(f"Whisper cannot process {subtitle.video.original_path} because of unsupported audio track language: '{subtitle.audio_language}'")
subtitle.content = None
return
logger.info(f'Starting WhisperAI {subtitle.task} to {language_from_alpha3(output_language)} for {subtitle.video.original_path}')
startTime = time.time()
r = self.session.post(f"{self.endpoint}/asr",
params={'task': subtitle.task, 'language': input_language, 'output': 'srt', 'encode': 'false'},
files={'audio_file': out},
timeout=(self.response, self.timeout))
endTime = time.time()
elapsedTime = timedelta(seconds=round(endTime - startTime))
# for debugging, log if anything got returned
subtitle_length = len(r.content)
logger.debug(f'Returned subtitle length is {subtitle_length:,} bytes')
subtitle_length = min(subtitle_length, 1000)
if subtitle_length > 0:
logger.debug(f'First {subtitle_length} bytes of subtitle: {r.content[0:subtitle_length]}')
logger.info(f'Completed WhisperAI {subtitle.task} to {language_from_alpha3(output_language)} in {elapsedTime} for {subtitle.video.original_path}')
subtitle.content = r.content