9664421460

Committed 25 Jun 2024 02:31PM UTC coverage: 70.938%. First build

Build # 9664421460

Build Type

Pull #1139

github

Committed by

web-flow

Commit Message

Merge 0834a594a into 46941e1c4

Pull Request Pull Request #1139: [WIP] Ptitloup/remote encoding fix

Run Details

3 of 16 new or added lines in 3 files covered. (18.75%)

12031 of 16960 relevant lines covered (70.94%)

0.71 hits per line

Source File
Press 'n' to go to next uncovered line, 'b' for previous

0.0

/pod/video_encode_transcript/transcript_model.py

import numpy as np
import shlex
import subprocess
import json

import sys
import os
from timeit import default_timer as timer
import datetime as dt
from datetime import timedelta

from webvtt import WebVTT, Caption
from shlex import quote

import logging

try:
    from ..custom import settings_local
except ImportError:
    from .. import settings as settings_local

from .encoding_utils import sec_to_timestamp

DEBUG = getattr(settings_local, "DEBUG", False)

TRANSCRIPTION_MODEL_PARAM = getattr(settings_local, "TRANSCRIPTION_MODEL_PARAM", False)
USE_TRANSCRIPTION = getattr(settings_local, "USE_TRANSCRIPTION", False)
if USE_TRANSCRIPTION:
    TRANSCRIPTION_TYPE = getattr(settings_local, "TRANSCRIPTION_TYPE", "VOSK")
    if TRANSCRIPTION_TYPE == "VOSK":
        from vosk import Model, KaldiRecognizer
    elif TRANSCRIPTION_TYPE == "STT":
        from stt import Model
    elif TRANSCRIPTION_TYPE == "WHISPER":
        import whisper

TRANSCRIPTION_NORMALIZE = getattr(settings_local, "TRANSCRIPTION_NORMALIZE", False)
TRANSCRIPTION_NORMALIZE_TARGET_LEVEL = getattr(
    settings_local, "TRANSCRIPTION_NORMALIZE_TARGET_LEVEL", -16.0
)

TRANSCRIPTION_AUDIO_SPLIT_TIME = getattr(
    settings_local, "TRANSCRIPTION_AUDIO_SPLIT_TIME", 600
)  # 10min
# time in sec for phrase length
TRANSCRIPTION_STT_SENTENCE_MAX_LENGTH = getattr(
    settings_local, "TRANSCRIPTION_STT_SENTENCE_MAX_LENGTH", 2
)
TRANSCRIPTION_STT_SENTENCE_BLANK_SPLIT_TIME = getattr(
    settings_local, "TRANSCRIPTION_STT_SENTENCE_BLANK_SPLIT_TIME", 0.5
)
log = logging.getLogger(__name__)


def get_model(lang):
    """Get model for STT or Vosk software to transcript audio."""
    transript_model = Model(TRANSCRIPTION_MODEL_PARAM[TRANSCRIPTION_TYPE][lang]["model"])
    if TRANSCRIPTION_TYPE == "STT":
        if TRANSCRIPTION_MODEL_PARAM[TRANSCRIPTION_TYPE][lang].get("beam_width"):
            transript_model.setBeamWidth(
                TRANSCRIPTION_MODEL_PARAM[TRANSCRIPTION_TYPE][lang]["beam_width"]
            )
        if TRANSCRIPTION_MODEL_PARAM[TRANSCRIPTION_TYPE][lang].get("scorer"):
            print(
                "Loading scorer from files {}".format(
                    TRANSCRIPTION_MODEL_PARAM[TRANSCRIPTION_TYPE][lang]["scorer"]
                ),
                file=sys.stderr,
            )
            scorer_load_start = timer()
            transript_model.enableExternalScorer(
                TRANSCRIPTION_MODEL_PARAM[TRANSCRIPTION_TYPE][lang]["scorer"]
            )
            scorer_load_end = timer() - scorer_load_start
            print("Loaded scorer in {:.3}s.".format(scorer_load_end), file=sys.stderr)
            if TRANSCRIPTION_MODEL_PARAM[TRANSCRIPTION_TYPE][lang].get(
                "lm_alpha"
            ) and TRANSCRIPTION_MODEL_PARAM[TRANSCRIPTION_TYPE][lang].get("lm_beta"):
                transript_model.setScorerAlphaBeta(
                    TRANSCRIPTION_MODEL_PARAM[TRANSCRIPTION_TYPE][lang]["lm_alpha"],
                    TRANSCRIPTION_MODEL_PARAM[TRANSCRIPTION_TYPE][lang]["lm_beta"],
                )
    return transript_model


def start_transcripting(mp3filepath, duration, lang):
    """
    Start direct transcription.

    Normalize the audio if set, get the model according to the lang and start transcript.
    """
    if TRANSCRIPTION_NORMALIZE:
        mp3filepath = normalize_mp3(mp3filepath)
    if TRANSCRIPTION_TYPE == "WHISPER":
        msg, webvtt, all_text = main_whisper_transcript(mp3filepath, duration, lang)
    else:
        transript_model = get_model(lang)
        msg, webvtt, all_text = start_main_transcript(
            mp3filepath, duration, transript_model
        )
    if DEBUG:
        print(msg)
        print(webvtt)
        print("\n%s\n" % all_text)

    return msg, webvtt


def start_main_transcript(mp3filepath, duration, transript_model):
    """Call transcription depending software type."""
    if TRANSCRIPTION_TYPE == "STT":
        msg, webvtt, all_text = main_stt_transcript(
            mp3filepath, duration, transript_model
        )
    elif TRANSCRIPTION_TYPE == "VOSK":
        msg, webvtt, all_text = main_vosk_transcript(
            mp3filepath, duration, transript_model
        )
    return msg, webvtt, all_text


def convert_samplerate(audio_path, desired_sample_rate, trim_start, duration):
    """Convert audio to subaudio and add good sample rate."""
    sox_cmd = "sox {} --type raw --bits 16 --channels 1 --rate {} ".format(
        quote(audio_path), desired_sample_rate
    )
    sox_cmd += "--encoding signed-integer --endian little --compression 0.0 "
    sox_cmd += "--no-dither - trim {} {}".format(trim_start, duration)

    try:
        output = subprocess.check_output(shlex.split(sox_cmd), stderr=subprocess.PIPE)

    except subprocess.CalledProcessError as e:
        raise RuntimeError("SoX returned non-zero status: {}".format(e.stderr))
    except OSError as e:
        raise OSError(
            e.errno,
            "SoX not found, use {}hz files or install it: {}".format(
                desired_sample_rate, e.strerror
            ),
        )

    if TRANSCRIPTION_TYPE == "WHISPER":
        return np.frombuffer(output, np.int16).flatten().astype(np.float32) / 32768.0
    else:
        return np.frombuffer(output, np.int16)


def normalize_mp3(mp3filepath):
    """Normalize the audio to good format and sound level."""
    filename, file_extension = os.path.splitext(mp3filepath)
    mp3normfile = "{}{}{}".format(filename, "_norm", file_extension)
    normalize_cmd = "ffmpeg-normalize {} ".format(quote(mp3filepath))
    normalize_cmd += "-c:a libmp3lame -b:a 192k --normalization-type ebu "
    # normalize_cmd += \
    # '--loudness-range-target 7.0 --true-peak 0.0 --offset 0.0 '
    normalize_cmd += "--target-level {} -f -o {}".format(
        TRANSCRIPTION_NORMALIZE_TARGET_LEVEL, quote(mp3normfile)
    )
    if DEBUG:
        print(normalize_cmd)
    try:
        subprocess.check_output(shlex.split(normalize_cmd), stderr=subprocess.PIPE)
        return mp3normfile
    except subprocess.CalledProcessError as e:
        log.error("ffmpeg-normalize returned non-zero status: {}".format(e.stderr))
        return mp3filepath
    except OSError as e:
        log.error("ffmpeg-normalize not found {}".format(e.strerror))
        return mp3filepath


# #################################
# TRANSCRIPT VIDEO: MAIN FUNCTION
# #################################


def convert_vosk_samplerate(audio_path, desired_sample_rate, trim_start, duration):
    """Convert audio to the good sample rate."""
    sox_cmd = "sox {} --type raw --bits 16 --channels 1 --rate {} ".format(
        quote(audio_path), desired_sample_rate
    )
    sox_cmd += "--encoding signed-integer --endian little --compression 0.0 "
    sox_cmd += "--no-dither - trim {} {}".format(trim_start, duration)

    try:
        output = subprocess.Popen(shlex.split(sox_cmd), stdout=subprocess.PIPE)

    except subprocess.CalledProcessError as e:
        raise RuntimeError("SoX returned non-zero status: {}".format(e.stderr))
    except OSError as e:
        raise OSError(
            e.errno,
            "SoX not found, use {}hz files or install it: {}".format(
                desired_sample_rate, e.strerror
            ),
        )
    return output


def get_word_result_from_data(results, audio, rec):
    """Get subsound from audio and add transcription to result parameter."""
    while True:
        data = audio.stdout.read(4000)
        if len(data) == 0:
            break
        if rec.AcceptWaveform(data):
            results.append(rec.Result())
    results.append(rec.Result())


def words_to_vtt(
    words,
    start_trim,
    duration,
    is_first_caption,
    text_caption,
    start_caption,
    last_word_added,
    all_text,
    webvtt,
):
    """Convert word and time to webvtt captions."""
    for index, word in enumerate(words):
        start_key = "start_time"
        word_duration = word.get("duration", 0)
        last_word = words[-1]
        last_word_duration = last_word.get("duration", 0)
        if TRANSCRIPTION_TYPE == "VOSK":
            start_key = "start"
            word_duration = word["end"] - word["start"]
            last_word_duration = words[-1]["end"] - words[-1]["start"]
        next_word = None
        blank_duration = 0
        if word != words[-1] and (index + 1) < len(words):
            next_word = words[index + 1]
            blank_duration = ((next_word[start_key]) - start_caption) - (
                ((word[start_key]) - start_caption) + word_duration
            )
        all_text += word["word"] + " "
        # word: <class 'dict'> {'word': 'bonjour', 'start ':
        # 0.58, 'duration': 7.34}
        text_caption.append(word["word"])
        if not (
            (((word[start_key]) - start_caption) < TRANSCRIPTION_STT_SENTENCE_MAX_LENGTH)
            and (
                next_word is not None
                and (blank_duration < TRANSCRIPTION_STT_SENTENCE_BLANK_SPLIT_TIME)
            )
        ):
            # on créé le caption
            if is_first_caption:
                # A revoir, fusion de la nouvelle ligne avec
                # l'ancienne...
                is_first_caption = False
                text_caption = get_text_caption(text_caption, last_word_added)

            stop_caption = word[start_key] + word_duration

            # on evite le chevauchement
            change_previous_end_caption(webvtt, start_caption)

            caption = Caption(
                sec_to_timestamp(start_caption),
                sec_to_timestamp(stop_caption),
                " ".join(text_caption),
            )

            webvtt.captions.append(caption)
            # on remet tout à zero pour la prochaine phrase
            start_caption = word[start_key]
            text_caption = []
            last_word_added = word["word"]
    if start_trim + TRANSCRIPTION_AUDIO_SPLIT_TIME > duration:
        # on ajoute ici la dernière phrase de la vidéo
        stop_caption = start_trim + words[-1][start_key] + last_word_duration
        caption = Caption(
            sec_to_timestamp(start_caption),
            sec_to_timestamp(stop_caption),
            " ".join(text_caption),
        )
        webvtt.captions.append(caption)
    return all_text, webvtt


def main_vosk_transcript(norm_mp3_file, duration, transript_model):
    """Vosk transcription."""
    msg = ""
    inference_start = timer()
    msg += "\nInference start %0.3fs." % inference_start
    desired_sample_rate = 16000

    rec = KaldiRecognizer(transript_model, desired_sample_rate)
    rec.SetWords(True)

    webvtt = WebVTT()
    all_text = ""
    for start_trim in range(0, duration, TRANSCRIPTION_AUDIO_SPLIT_TIME):
        audio = convert_vosk_samplerate(
            norm_mp3_file,
            desired_sample_rate,
            start_trim,
            TRANSCRIPTION_AUDIO_SPLIT_TIME,  # dur
        )
        msg += "\nRunning inference."
        results = []
        get_word_result_from_data(results, audio, rec)
        for res in results:
            words = json.loads(res).get("result")
            text = json.loads(res).get("text")
            if not words:
                continue
            start_caption = words[0]["start"]
            stop_caption = words[-1]["end"]
            caption = Caption(
                sec_to_timestamp(start_caption),
                sec_to_timestamp(stop_caption),
                text,
            )
            webvtt.captions.append(caption)
            """
            text_caption = []
            is_first_caption = True
            all_text, webvtt = words_to_vtt(
                words,
                start_trim,
                duration,
                is_first_caption,
                text_caption,
                start_caption,
                last_word_added,
                all_text,
                webvtt,
            )
            """
    inference_end = timer() - inference_start

    msg += "\nInference took %0.3fs." % inference_end
    return msg, webvtt, all_text


def main_stt_transcript(norm_mp3_file, duration, transript_model):
    """STT transcription."""
    msg = ""
    inference_start = timer()
    msg += "\nInference start %0.3fs." % inference_start
    desired_sample_rate = transript_model.sampleRate()
    webvtt = WebVTT()
    last_word_added = ""
    metadata = None
    all_text = ""
    for start_trim in range(0, duration, TRANSCRIPTION_AUDIO_SPLIT_TIME):
        end_trim = (
            duration
            if start_trim + TRANSCRIPTION_AUDIO_SPLIT_TIME > duration
            else (
                start_trim
                + TRANSCRIPTION_AUDIO_SPLIT_TIME
                + TRANSCRIPTION_STT_SENTENCE_MAX_LENGTH
            )
        )

        dur = (
            (TRANSCRIPTION_AUDIO_SPLIT_TIME + TRANSCRIPTION_STT_SENTENCE_MAX_LENGTH)
            if (
                (
                    start_trim
                    + TRANSCRIPTION_AUDIO_SPLIT_TIME
                    + TRANSCRIPTION_STT_SENTENCE_MAX_LENGTH
                )
                < duration
            )
            else (duration - start_trim)
        )

        msg += "\ntake audio from %s to %s - %s" % (start_trim, end_trim, dur)

        audio = convert_samplerate(norm_mp3_file, desired_sample_rate, start_trim, dur)
        msg += "\nRunning inference."

        metadata = transript_model.sttWithMetadata(audio)

        for transcript in metadata.transcripts:
            msg += "\nConfidence: %s" % transcript.confidence
            words = words_from_candidate_transcript(transcript)
            start_caption = start_trim + words[0]["start_time"]
            text_caption = []
            is_first_caption = True
            all_text, webvtt = words_to_vtt(
                words,
                start_trim,
                duration,
                is_first_caption,
                text_caption,
                start_caption,
                last_word_added,
                all_text,
                webvtt,
            )
    inference_end = timer() - inference_start

    msg += "\nInference took %0.3fs." % inference_end
    return msg, webvtt, all_text


def main_whisper_transcript(norm_mp3_file, duration, lang):
    """Whisper transcription."""
    msg = ""
    all_text = ""
    webvtt = WebVTT()
    inference_start = timer()
    desired_sample_rate = 16000
    msg += "\nInference start %0.3fs." % inference_start

    model = whisper.load_model(
        TRANSCRIPTION_MODEL_PARAM[TRANSCRIPTION_TYPE][lang]["model"],
        download_root=TRANSCRIPTION_MODEL_PARAM[TRANSCRIPTION_TYPE][lang][
            "download_root"
        ],
    )
    '''
    for start_trim in range(0, duration, TRANSCRIPTION_AUDIO_SPLIT_TIME):
        log.info("start_trim: " + str(start_trim))
        audio = convert_samplerate(
            norm_mp3_file,
            desired_sample_rate,
            start_trim,
            TRANSCRIPTION_AUDIO_SPLIT_TIME,  # dur
        )
        transcription = model.transcribe(audio, language=lang)
        msg += "\nRunning inference."
        for segment in transcription["segments"]:
            caption = Caption(
                sec_to_timestamp(segment["start"] + start_trim),
                sec_to_timestamp(segment["end"] + start_trim),
                segment["text"],
            )
            webvtt.captions.append(caption)
    '''
    audio = convert_samplerate(
        norm_mp3_file,
        desired_sample_rate,
        0,
        duration
    )
    transcription = model.transcribe(audio, language=lang)
    for segment in transcription["segments"]:
        caption = Caption(
            sec_to_timestamp(segment["start"]),
            sec_to_timestamp(segment["end"]),
            segment["text"],
        )
        webvtt.captions.append(caption)
    inference_end = timer() - inference_start
    msg += "\nInference took %0.3fs." % inference_end
    return msg, webvtt, all_text


def change_previous_end_caption(webvtt, start_caption):
    """Change the end time for caption."""
    if len(webvtt.captions) > 0:
        prev_end = dt.datetime.strptime(webvtt.captions[-1].end, "%H:%M:%S.%f")
        td_prev_end = timedelta(
            hours=prev_end.hour,
            minutes=prev_end.minute,
            seconds=prev_end.second,
            microseconds=prev_end.microsecond,
        ).total_seconds()
        if td_prev_end > start_caption:
            webvtt.captions[-1].end = sec_to_timestamp(start_caption)


def get_text_caption(text_caption, last_word_added):
    """Get the text for a caption."""
    try:
        first_index = text_caption.index(last_word_added)
        return text_caption[first_index + 1 :]
    except ValueError:
        return text_caption


def words_from_candidate_transcript(metadata):
    """Get words list from transcription."""
    word = ""
    word_list = []
    word_start_time = 0
    # Loop through each character
    for i, token in enumerate(metadata.tokens):
        # Append character to word if it's not a space
        if token.text != " ":
            if len(word) == 0:
                # Log the start time of the new word
                word_start_time = token.start_time

            word = word + token.text
        # Word boundary is either a space or the last character in the array
        if token.text == " " or i == len(metadata.tokens) - 1:
            word_duration = token.start_time - word_start_time

            if word_duration < 0:
                word_duration = 0

            each_word = dict()
            each_word["word"] = word
            each_word["start_time"] = round(word_start_time, 4)
            each_word["duration"] = round(word_duration, 4)

            word_list.append(each_word)
            # Reset
            word = ""
            word_start_time = 0

    return word_list

1	import numpy as np	×
2	import shlex	×
3	import subprocess	×
4	import json	×
5
6	import sys	×
7	import os	×
8	from timeit import default_timer as timer	×
9	import datetime as dt	×
10	from datetime import timedelta	×
11
12	from webvtt import WebVTT, Caption	×
13	from shlex import quote	×
14
15	import logging	×
16
17	try:	×
18	from ..custom import settings_local	×
19	except ImportError:	×
20	from .. import settings as settings_local	×
21
22	from .encoding_utils import sec_to_timestamp	×
23
24	DEBUG = getattr(settings_local, "DEBUG", False)	×
25
26	TRANSCRIPTION_MODEL_PARAM = getattr(settings_local, "TRANSCRIPTION_MODEL_PARAM", False)	×
27	USE_TRANSCRIPTION = getattr(settings_local, "USE_TRANSCRIPTION", False)	×
28	if USE_TRANSCRIPTION:	×
29	TRANSCRIPTION_TYPE = getattr(settings_local, "TRANSCRIPTION_TYPE", "VOSK")	×
30	if TRANSCRIPTION_TYPE == "VOSK":	×
31	from vosk import Model, KaldiRecognizer	×
32	elif TRANSCRIPTION_TYPE == "STT":	×
33	from stt import Model	×
34	elif TRANSCRIPTION_TYPE == "WHISPER":	×
35	import whisper	×
36
37	TRANSCRIPTION_NORMALIZE = getattr(settings_local, "TRANSCRIPTION_NORMALIZE", False)	×
38	TRANSCRIPTION_NORMALIZE_TARGET_LEVEL = getattr(	×
39	settings_local, "TRANSCRIPTION_NORMALIZE_TARGET_LEVEL", -16.0
40	)
41
42	TRANSCRIPTION_AUDIO_SPLIT_TIME = getattr(	×
43	settings_local, "TRANSCRIPTION_AUDIO_SPLIT_TIME", 600
44	) # 10min
45	# time in sec for phrase length
46	TRANSCRIPTION_STT_SENTENCE_MAX_LENGTH = getattr(	×
47	settings_local, "TRANSCRIPTION_STT_SENTENCE_MAX_LENGTH", 2
48	)
49	TRANSCRIPTION_STT_SENTENCE_BLANK_SPLIT_TIME = getattr(	×
50	settings_local, "TRANSCRIPTION_STT_SENTENCE_BLANK_SPLIT_TIME", 0.5
51	)
52	log = logging.getLogger(__name__)	×
53
54
55	def get_model(lang):	×
56	"""Get model for STT or Vosk software to transcript audio."""
57	transript_model = Model(TRANSCRIPTION_MODEL_PARAM[TRANSCRIPTION_TYPE][lang]["model"])	×
58	if TRANSCRIPTION_TYPE == "STT":	×
59	if TRANSCRIPTION_MODEL_PARAM[TRANSCRIPTION_TYPE][lang].get("beam_width"):	×
60	transript_model.setBeamWidth(	×
61	TRANSCRIPTION_MODEL_PARAM[TRANSCRIPTION_TYPE][lang]["beam_width"]
62	)
63	if TRANSCRIPTION_MODEL_PARAM[TRANSCRIPTION_TYPE][lang].get("scorer"):	×
64	print(	×
65	"Loading scorer from files {}".format(
66	TRANSCRIPTION_MODEL_PARAM[TRANSCRIPTION_TYPE][lang]["scorer"]
67	),
68	file=sys.stderr,
69	)
70	scorer_load_start = timer()	×
71	transript_model.enableExternalScorer(	×
72	TRANSCRIPTION_MODEL_PARAM[TRANSCRIPTION_TYPE][lang]["scorer"]
73	)
74	scorer_load_end = timer() - scorer_load_start	×
75	print("Loaded scorer in {:.3}s.".format(scorer_load_end), file=sys.stderr)	×
76	if TRANSCRIPTION_MODEL_PARAM[TRANSCRIPTION_TYPE][lang].get(	×
77	"lm_alpha"
78	) and TRANSCRIPTION_MODEL_PARAM[TRANSCRIPTION_TYPE][lang].get("lm_beta"):
79	transript_model.setScorerAlphaBeta(	×
80	TRANSCRIPTION_MODEL_PARAM[TRANSCRIPTION_TYPE][lang]["lm_alpha"],
81	TRANSCRIPTION_MODEL_PARAM[TRANSCRIPTION_TYPE][lang]["lm_beta"],
82	)
83	return transript_model	×
84
85
86	def start_transcripting(mp3filepath, duration, lang):	×
87	"""
88	Start direct transcription.
89
90	Normalize the audio if set, get the model according to the lang and start transcript.
91	"""
92	if TRANSCRIPTION_NORMALIZE:	×
93	mp3filepath = normalize_mp3(mp3filepath)	×
94	if TRANSCRIPTION_TYPE == "WHISPER":	×
95	msg, webvtt, all_text = main_whisper_transcript(mp3filepath, duration, lang)	×
96	else:
97	transript_model = get_model(lang)	×
98	msg, webvtt, all_text = start_main_transcript(	×
99	mp3filepath, duration, transript_model
100	)
101	if DEBUG:	×
102	print(msg)	×
103	print(webvtt)	×
104	print("\n%s\n" % all_text)	×
105
106	return msg, webvtt	×
107
108
109	def start_main_transcript(mp3filepath, duration, transript_model):	×
110	"""Call transcription depending software type."""
111	if TRANSCRIPTION_TYPE == "STT":	×
112	msg, webvtt, all_text = main_stt_transcript(	×
113	mp3filepath, duration, transript_model
114	)
115	elif TRANSCRIPTION_TYPE == "VOSK":	×
116	msg, webvtt, all_text = main_vosk_transcript(	×
117	mp3filepath, duration, transript_model
118	)
119	return msg, webvtt, all_text	×
120
121
122	def convert_samplerate(audio_path, desired_sample_rate, trim_start, duration):	×
123	"""Convert audio to subaudio and add good sample rate."""
124	sox_cmd = "sox {} --type raw --bits 16 --channels 1 --rate {} ".format(	×
125	quote(audio_path), desired_sample_rate
126	)
127	sox_cmd += "--encoding signed-integer --endian little --compression 0.0 "	×
128	sox_cmd += "--no-dither - trim {} {}".format(trim_start, duration)	×
129
130	try:	×
131	output = subprocess.check_output(shlex.split(sox_cmd), stderr=subprocess.PIPE)	×
132
133	except subprocess.CalledProcessError as e:	×
134	raise RuntimeError("SoX returned non-zero status: {}".format(e.stderr))	×
135	except OSError as e:	×
136	raise OSError(	×
137	e.errno,
138	"SoX not found, use {}hz files or install it: {}".format(
139	desired_sample_rate, e.strerror
140	),
141	)
142
143	if TRANSCRIPTION_TYPE == "WHISPER":	×
144	return np.frombuffer(output, np.int16).flatten().astype(np.float32) / 32768.0	×
145	else:
146	return np.frombuffer(output, np.int16)	×
147
148
149	def normalize_mp3(mp3filepath):	×
150	"""Normalize the audio to good format and sound level."""
151	filename, file_extension = os.path.splitext(mp3filepath)	×
152	mp3normfile = "{}{}{}".format(filename, "_norm", file_extension)	×
153	normalize_cmd = "ffmpeg-normalize {} ".format(quote(mp3filepath))	×
154	normalize_cmd += "-c:a libmp3lame -b:a 192k --normalization-type ebu "	×
155	# normalize_cmd += \
156	# '--loudness-range-target 7.0 --true-peak 0.0 --offset 0.0 '
157	normalize_cmd += "--target-level {} -f -o {}".format(	×
158	TRANSCRIPTION_NORMALIZE_TARGET_LEVEL, quote(mp3normfile)
159	)
160	if DEBUG:	×
161	print(normalize_cmd)	×
162	try:	×
163	subprocess.check_output(shlex.split(normalize_cmd), stderr=subprocess.PIPE)	×
164	return mp3normfile	×
165	except subprocess.CalledProcessError as e:	×
166	log.error("ffmpeg-normalize returned non-zero status: {}".format(e.stderr))	×
167	return mp3filepath	×
168	except OSError as e:	×
169	log.error("ffmpeg-normalize not found {}".format(e.strerror))	×
170	return mp3filepath	×
171
172
173	# #################################
174	# TRANSCRIPT VIDEO: MAIN FUNCTION
175	# #################################
176
177
178	def convert_vosk_samplerate(audio_path, desired_sample_rate, trim_start, duration):	×
179	"""Convert audio to the good sample rate."""
180	sox_cmd = "sox {} --type raw --bits 16 --channels 1 --rate {} ".format(	×
181	quote(audio_path), desired_sample_rate
182	)
183	sox_cmd += "--encoding signed-integer --endian little --compression 0.0 "	×
184	sox_cmd += "--no-dither - trim {} {}".format(trim_start, duration)	×
185
186	try:	×
187	output = subprocess.Popen(shlex.split(sox_cmd), stdout=subprocess.PIPE)	×
188
189	except subprocess.CalledProcessError as e:	×
190	raise RuntimeError("SoX returned non-zero status: {}".format(e.stderr))	×
191	except OSError as e:	×
192	raise OSError(	×
193	e.errno,
194	"SoX not found, use {}hz files or install it: {}".format(
195	desired_sample_rate, e.strerror
196	),
197	)
198	return output	×
199
200
201	def get_word_result_from_data(results, audio, rec):	×
202	"""Get subsound from audio and add transcription to result parameter."""
203	while True:
204	data = audio.stdout.read(4000)	×
205	if len(data) == 0:	×
206	break	×
207	if rec.AcceptWaveform(data):	×
208	results.append(rec.Result())	×
209	results.append(rec.Result())	×
210
211
212	def words_to_vtt(	×
213	words,
214	start_trim,
215	duration,
216	is_first_caption,
217	text_caption,
218	start_caption,
219	last_word_added,
220	all_text,
221	webvtt,
222	):
223	"""Convert word and time to webvtt captions."""
224	for index, word in enumerate(words):	×
225	start_key = "start_time"	×
226	word_duration = word.get("duration", 0)	×
227	last_word = words[-1]	×
228	last_word_duration = last_word.get("duration", 0)	×
229	if TRANSCRIPTION_TYPE == "VOSK":	×
230	start_key = "start"	×
231	word_duration = word["end"] - word["start"]	×
232	last_word_duration = words[-1]["end"] - words[-1]["start"]	×
233	next_word = None	×
234	blank_duration = 0	×
235	if word != words[-1] and (index + 1) < len(words):	×
236	next_word = words[index + 1]	×
237	blank_duration = ((next_word[start_key]) - start_caption) - (	×
238	((word[start_key]) - start_caption) + word_duration
239	)
240	all_text += word["word"] + " "	×
241	# word: <class 'dict'> {'word': 'bonjour', 'start ':
242	# 0.58, 'duration': 7.34}
243	text_caption.append(word["word"])	×
244	if not (	×
245	(((word[start_key]) - start_caption) < TRANSCRIPTION_STT_SENTENCE_MAX_LENGTH)
246	and (
247	next_word is not None
248	and (blank_duration < TRANSCRIPTION_STT_SENTENCE_BLANK_SPLIT_TIME)
249	)
250	):
251	# on créé le caption
252	if is_first_caption:	×
253	# A revoir, fusion de la nouvelle ligne avec
254	# l'ancienne...
255	is_first_caption = False	×
256	text_caption = get_text_caption(text_caption, last_word_added)	×
257
258	stop_caption = word[start_key] + word_duration	×
259
260	# on evite le chevauchement
261	change_previous_end_caption(webvtt, start_caption)	×
262
263	caption = Caption(	×
264	sec_to_timestamp(start_caption),
265	sec_to_timestamp(stop_caption),
266	" ".join(text_caption),
267	)
268
269	webvtt.captions.append(caption)	×
270	# on remet tout à zero pour la prochaine phrase
271	start_caption = word[start_key]	×
272	text_caption = []	×
273	last_word_added = word["word"]	×
274	if start_trim + TRANSCRIPTION_AUDIO_SPLIT_TIME > duration:	×
275	# on ajoute ici la dernière phrase de la vidéo
276	stop_caption = start_trim + words[-1][start_key] + last_word_duration	×
277	caption = Caption(	×
278	sec_to_timestamp(start_caption),
279	sec_to_timestamp(stop_caption),
280	" ".join(text_caption),
281	)
282	webvtt.captions.append(caption)	×
283	return all_text, webvtt	×
284
285
286	def main_vosk_transcript(norm_mp3_file, duration, transript_model):	×
287	"""Vosk transcription."""
288	msg = ""	×
289	inference_start = timer()	×
290	msg += "\nInference start %0.3fs." % inference_start	×
291	desired_sample_rate = 16000	×
292
293	rec = KaldiRecognizer(transript_model, desired_sample_rate)	×
294	rec.SetWords(True)	×
295
296	webvtt = WebVTT()	×
297	all_text = ""	×
298	for start_trim in range(0, duration, TRANSCRIPTION_AUDIO_SPLIT_TIME):	×
299	audio = convert_vosk_samplerate(	×
300	norm_mp3_file,
301	desired_sample_rate,
302	start_trim,
303	TRANSCRIPTION_AUDIO_SPLIT_TIME, # dur
304	)
305	msg += "\nRunning inference."	×
306	results = []	×
307	get_word_result_from_data(results, audio, rec)	×
308	for res in results:	×
309	words = json.loads(res).get("result")	×
310	text = json.loads(res).get("text")	×
311	if not words:	×
312	continue	×
313	start_caption = words[0]["start"]	×
314	stop_caption = words[-1]["end"]	×
315	caption = Caption(	×
316	sec_to_timestamp(start_caption),
317	sec_to_timestamp(stop_caption),
318	text,
319	)
320	webvtt.captions.append(caption)	×
321	"""	×
322	text_caption = []
323	is_first_caption = True
324	all_text, webvtt = words_to_vtt(
325	words,
326	start_trim,
327	duration,
328	is_first_caption,
329	text_caption,
330	start_caption,
331	last_word_added,
332	all_text,
333	webvtt,
334	)
335	"""
336	inference_end = timer() - inference_start	×
337
338	msg += "\nInference took %0.3fs." % inference_end	×
339	return msg, webvtt, all_text	×
340
341
342	def main_stt_transcript(norm_mp3_file, duration, transript_model):	×
343	"""STT transcription."""
344	msg = ""	×
345	inference_start = timer()	×
346	msg += "\nInference start %0.3fs." % inference_start	×
347	desired_sample_rate = transript_model.sampleRate()	×
348	webvtt = WebVTT()	×
349	last_word_added = ""	×
350	metadata = None	×
351	all_text = ""	×
352	for start_trim in range(0, duration, TRANSCRIPTION_AUDIO_SPLIT_TIME):	×
353	end_trim = (	×
354	duration
355	if start_trim + TRANSCRIPTION_AUDIO_SPLIT_TIME > duration
356	else (
357	start_trim
358	+ TRANSCRIPTION_AUDIO_SPLIT_TIME
359	+ TRANSCRIPTION_STT_SENTENCE_MAX_LENGTH
360	)
361	)
362
363	dur = (	×
364	(TRANSCRIPTION_AUDIO_SPLIT_TIME + TRANSCRIPTION_STT_SENTENCE_MAX_LENGTH)
365	if (
366	(
367	start_trim
368	+ TRANSCRIPTION_AUDIO_SPLIT_TIME
369	+ TRANSCRIPTION_STT_SENTENCE_MAX_LENGTH
370	)
371	< duration
372	)
373	else (duration - start_trim)
374	)
375
376	msg += "\ntake audio from %s to %s - %s" % (start_trim, end_trim, dur)	×
377
378	audio = convert_samplerate(norm_mp3_file, desired_sample_rate, start_trim, dur)	×
379	msg += "\nRunning inference."	×
380
381	metadata = transript_model.sttWithMetadata(audio)	×
382
383	for transcript in metadata.transcripts:	×
384	msg += "\nConfidence: %s" % transcript.confidence	×
385	words = words_from_candidate_transcript(transcript)	×
386	start_caption = start_trim + words[0]["start_time"]	×
387	text_caption = []	×
388	is_first_caption = True	×
389	all_text, webvtt = words_to_vtt(	×
390	words,
391	start_trim,
392	duration,
393	is_first_caption,
394	text_caption,
395	start_caption,
396	last_word_added,
397	all_text,
398	webvtt,
399	)
400	inference_end = timer() - inference_start	×
401
402	msg += "\nInference took %0.3fs." % inference_end	×
403	return msg, webvtt, all_text	×
404
405
406	def main_whisper_transcript(norm_mp3_file, duration, lang):	×
407	"""Whisper transcription."""
408	msg = ""	×
409	all_text = ""	×
410	webvtt = WebVTT()	×
411	inference_start = timer()	×
412	desired_sample_rate = 16000	×
413	msg += "\nInference start %0.3fs." % inference_start	×
414
415	model = whisper.load_model(	×
416	TRANSCRIPTION_MODEL_PARAM[TRANSCRIPTION_TYPE][lang]["model"],
417	download_root=TRANSCRIPTION_MODEL_PARAM[TRANSCRIPTION_TYPE][lang][
418	"download_root"
419	],
420	)
421	'''
422	for start_trim in range(0, duration, TRANSCRIPTION_AUDIO_SPLIT_TIME):
423	log.info("start_trim: " + str(start_trim))
424	audio = convert_samplerate(
425	norm_mp3_file,
426	desired_sample_rate,
427	start_trim,
428	TRANSCRIPTION_AUDIO_SPLIT_TIME, # dur
429	)
430	transcription = model.transcribe(audio, language=lang)
431	msg += "\nRunning inference."
432	for segment in transcription["segments"]:
433	caption = Caption(
434	sec_to_timestamp(segment["start"] + start_trim),
435	sec_to_timestamp(segment["end"] + start_trim),
436	segment["text"],
437	)
438	webvtt.captions.append(caption)
439	'''
NEW 440	audio = convert_samplerate(	×
441	norm_mp3_file,
442	desired_sample_rate,
443	0,
444	duration
445	)
NEW 446	transcription = model.transcribe(audio, language=lang)	×
NEW 447	for segment in transcription["segments"]:	×
NEW 448	caption = Caption(	×
449	sec_to_timestamp(segment["start"]),
450	sec_to_timestamp(segment["end"]),
451	segment["text"],
452	)
NEW 453	webvtt.captions.append(caption)	×
454	inference_end = timer() - inference_start	×
455	msg += "\nInference took %0.3fs." % inference_end	×
456	return msg, webvtt, all_text	×
457
458
459	def change_previous_end_caption(webvtt, start_caption):	×
460	"""Change the end time for caption."""
461	if len(webvtt.captions) > 0:	×
462	prev_end = dt.datetime.strptime(webvtt.captions[-1].end, "%H:%M:%S.%f")	×
463	td_prev_end = timedelta(	×
464	hours=prev_end.hour,
465	minutes=prev_end.minute,
466	seconds=prev_end.second,
467	microseconds=prev_end.microsecond,
468	).total_seconds()
469	if td_prev_end > start_caption:	×
470	webvtt.captions[-1].end = sec_to_timestamp(start_caption)	×
471
472
473	def get_text_caption(text_caption, last_word_added):	×
474	"""Get the text for a caption."""
475	try:	×
476	first_index = text_caption.index(last_word_added)	×
477	return text_caption[first_index + 1 :]	×
478	except ValueError:	×
479	return text_caption	×
480
481
482	def words_from_candidate_transcript(metadata):	×
483	"""Get words list from transcription."""
484	word = ""	×
485	word_list = []	×
486	word_start_time = 0	×
487	# Loop through each character
488	for i, token in enumerate(metadata.tokens):	×
489	# Append character to word if it's not a space
490	if token.text != " ":	×
491	if len(word) == 0:	×
492	# Log the start time of the new word
493	word_start_time = token.start_time	×
494
495	word = word + token.text	×
496	# Word boundary is either a space or the last character in the array
497	if token.text == " " or i == len(metadata.tokens) - 1:	×
498	word_duration = token.start_time - word_start_time	×
499
500	if word_duration < 0:	×
501	word_duration = 0	×
502
503	each_word = dict()	×
504	each_word["word"] = word	×
505	each_word["start_time"] = round(word_start_time, 4)	×
506	each_word["duration"] = round(word_duration, 4)	×
507
508	word_list.append(each_word)	×
509	# Reset
510	word = ""	×
511	word_start_time = 0	×
512
513	return word_list	×

EsupPortail / Esup-Pod / 9664421460

Source File Press 'n' to go to next uncovered line, 'b' for previous

Source File
Press 'n' to go to next uncovered line, 'b' for previous