j: ddlZddlZddlZddlZddlZddlmZmZddlm Z ddl m Z ddl m Z mZmZmZmZmZddlmZddlZddlZddlZddlmZddlmZmZdd lmZdd lm Z m!Z!dd l"m#Z#m$Z$m%Z%m&Z&dd l'm(Z(m)Z)m*Z*m+Z+eGd dZ,eGddZ-eGddZ.eGddZ/GddZ0GddZ1dee-dee2de3dee-fdZ4dej5dej6fdZ7d e8de9fd!Z:d"e!d#ee3deee3fd$Z;d%ee2d&e8d'e8ddfd(ZeZdZUeed<eed<eed<eed<dZdS)Wordstartendword probabilitycLtdtdt|S)NzIWord._asdict() method is deprecated, use dataclasses.asdict(Word) insteadr DeprecationWarningrselfs |/home/lighthouse/badge-speaker-benchmark/.benchmark-venv-community/lib/python3.11/site-packages/faster_whisper/transcribe.py_asdictz Word._asdict&s+ W     d||N)__name__ __module__ __qualname__float__annotations__strr)r*r(rrsL LLL JJJ IIIr*rceZdZUeed<eed<eed<eed<eed<eeed<eed<eed<eed <eee ed <eeed <d Z d S)Segmentidseekrrtexttokens avg_logprobcompression_rationo_speech_probwords temperaturecLtdtdt|S)NzOSegment._asdict() method is deprecated, use dataclasses.asdict(Segment) insteadr#r$r&s r(r)zSegment._asdict=s+ ]     d||r*N) r+r,r-intr/r.r0r r rr)r1r*r(r3r3/s GGG III LLL JJJ III I DJ %   r*r3ceZdZUeed<eed<eed<eed<eed<eed<eeed<eeed<eeed <eed <eed <eeed <ee e e efed <ee ed<eed<eeeed<eed<eed<eed<e ed<e ed<eed<eeed<e e eefed<eeed<ee ed<dS)TranscriptionOptions beam_sizebest_ofpatiencelength_penaltyrepetition_penaltyno_repeat_ngram_sizelog_prob_thresholdno_speech_thresholdcompression_ratio_thresholdcondition_on_previous_textprompt_reset_on_temperature temperaturesinitial_promptprefixsuppress_blanksuppress_tokenswithout_timestampsmax_initial_timestampword_timestampsprepend_punctuationsappend_punctuations multilingualmax_new_tokensclip_timestampshallucination_silence_thresholdhotwordsN) r+r,r-r>r/r.r boolr r r0rr1r*r(r@r@FsNNN LLLOOO '''!%(((!)%000 $$$$!&&&&u+U3 #567777 SMd3i((((    SM!!!3U +,,,,%-e_444smr*r@c~eZdZUeed<eed<eed<eed<eeeeefed<e ed<e ed<dS) TranscriptionInfolanguagelanguage_probabilitydurationduration_after_vadall_language_probstranscription_options vad_optionsN) r+r,r-r0r/r.r r r r@rr1r*r(r]r]dssMMMOOO eCJ&7!89999////r*r]cMzeZdZdZdZdejdedefdZ ddd d d d d d d gd dddddddddgddd ddd dddddddddd f#de e e ejfde e de dededededed ed!ed"e eeeeed#ffd$e ed%e ed&e ed'ed(ed)e e e eefd*e e d+ed,e eed-ed.ed/ed0e d1e d2ed3ed4e e eefd5e ed6e ed7e eed8e ed9ed:e e d;e edZd?ZdS)@BatchedInferencePipelinec"||_d|_dS)N)modellast_speech_timestamp)r'ris r(__init__z!BatchedInferencePipeline.__init__ps$) %("""r*c  ||\}}g}g}t||D]\d} tt| jjz} || jdd| | d\} } } |fd| D|jr4j ||||j |j j _ |S)Nr`r7offsetr tokenizerr7 time_offset segment_sizesegment_durationr5c8g|]}t|ddd|d|d|dt|dtdjjzS)r7r8r:rrrm)r6r8r:r7rrr9r5)dictdecodeget_compression_ratior>riframes_per_second).0 subsegmentchunk_metadataoutputr'ros r( z4BatchedInferencePipeline.forward..s#&--j.BCC$*=$9'-.>'?)(3(1&u-*?%,,Z-ABB++!*84tz7SS   r*) generate_segment_batchedzipr>rrirwappend_split_segments_by_timestampsrSadd_word_timestampsrTrUrj)r'featuresrochunks_metadataoptionsencoder_outputoutputssegmented_outputs segment_sizesr`rq subsegmentsr5single_timestamp_endingrzr{s` ` @@r(forwardz BatchedInferencePipeline.forwardwsv"&"?"? i# #  &)/7&C&C" "  "NF%j1HtH~~ 0LLMML   . . .  88#h'*84)!) 9  '  $ $'2    &  " )-)G)G!,+***D &! r*rrorc|jd}|j|j|jng|j|j|jt|jz}n |jj }||jj krPtdtd|tz d|d|jj d|jj d |j|}fd t|D}|j refd |jj |D}j} t!|D]\} } | || | <|jj|||j|j|j||j|jd d |jd|j|j } g} | D]o}t|jd}|jd||jzz}| t;||d zz |j|jdp|| fS)Nr)previous_tokensrQrZThe length of the prompt is , and the `max_new_tokens` C. Thus, the combined length of the prompt and `max_new_tokens` is: 6. This exceeds the `max_length` of the Whisper model: . You should either reduce the length of your prompt, or reduce the value of `max_new_tokens`, so that their combined length is less that .c8g|]}Sr1)copy)rx_prompts r(r|zEBatchedInferencePipeline.generate_segment_batched..s!<<.sD!#// a0@0CDDr*T) rArCrD max_lengthrOrP return_scoresreturn_no_speech_probsampling_temperaturerErF)r8r:r7)shaperi get_promptrMencoderQrZrWlenr ValueErrorrangerVdetect_languageindexr^ enumerategeneraterArCrDrOrPrLrErF sequences_idsscoresrrtr:)r'rror batch_sizerrpromptslanguage_tokenslanguage_token_indexilanguage_tokenresultsr{resultseq_len cum_logprobrs ` @r(r}z1BatchedInferencePipeline.generate_segment_batcheds ^A& && )5  !7888&9%'     ! -Vw'==JJ.J  - - -Ws6{{WWF +WW,6WW7;j6KWW ?Cj>S WWW **844<<<<% *;*;<<<   B%)Z%5%E%En%U%UO$*<< 0B#C#C %.%?%? B B!>3A /00*"++  '%"1!"1#3"&!(!5a!8&9!(!=,     F&.q122G -*gw7M.MNK MM +w{ ;#)#8!/2    v%%r*N transcribeFrrrhg?g?333333?g??333333@rT?r "'“¿([{-"'.。,，!！?？:：”)]}、audior^task log_progressrArBrCrDrErFr<.rIrGrHrJrKrMrNrOrPrQrRrSrTrUrV vad_filtervad_parametersrW chunk_lengthrXrYrrZlanguage_detection_thresholdlanguage_detection_segmentsreturnc%  4jjj4|r2jjjs!jjdd}t |tjst|4}|j d4z }%jj dt|%|pjjj }|s|ru|t|d}nPt |tr;d |vr|d td?i|d |i}t%||}n'|%|krd|j dd g}nt'd d}&t)||| \}'}(nd }&4fd|D}gg}(}'t+|D]\})}*|'||*d|*d|*d|*dz 4z }+|+dkr jjd|)|(|*d4z |+|*gdt/d|D4z },jj dt|%|,z |,rfd|'Dng}-d}.|jjjsd}d}/njtj|-tjjjjdfddgzd|$|#\}}/}.jj d||/n=jjjs*|dkr$jjd|zd}d}/t9jjjjj|| }0|-rtjd!|-Dng}-t?d?id"|d#|d$|d%|d&| d'| d(| d)|d*| d+t | t@tBfr | ddn| gd,|d-|d.|d/|rtE|0|n|d0|d1|d2|d3|"d4|d5dd6dd7|d8d9d:|d;|d}2$|-|0|(|!|1|}3|&stK|3|4}3|3|2fS)@attranscribe audio in chunks in batched fashion and return with language info. Arguments: audio: Path to the input file (or a file-like object), or the audio waveform. language: The language spoken in the audio. It should be a language code such as "en" or "fr". If not set, the language will be detected in the first 30 seconds of audio. task: Task to execute (transcribe or translate). log_progress: whether to show progress bar or not. beam_size: Beam size to use for decoding. best_of: Number of candidates when sampling with non-zero temperature. patience: Beam search patience factor. length_penalty: Exponential length penalty constant. repetition_penalty: Penalty applied to the score of previously generated tokens (set > 1 to penalize). no_repeat_ngram_size: Prevent repetitions of ngrams with this size (set 0 to disable). temperature: Temperature for sampling. If a list or tuple is passed, only the first value is used. initial_prompt: Optional text string or iterable of token ids to provide as a prompt for the each window. suppress_blank: Suppress blank outputs at the beginning of the sampling. suppress_tokens: List of token IDs to suppress. -1 will suppress a default set of symbols as defined in `tokenizer.non_speech_tokens()`. without_timestamps: Only sample text tokens. word_timestamps: Extract word-level timestamps using the cross-attention pattern and dynamic time warping, and include the timestamps for each word in each segment. Set as False. prepend_punctuations: If word_timestamps is True, merge these punctuation symbols with the next word append_punctuations: If word_timestamps is True, merge these punctuation symbols with the previous word multilingual: Perform language detection on every segment. vad_filter: Enable the voice activity detection (VAD) to filter out parts of the audio without speech. This step is using the Silero VAD model https://github.com/snakers4/silero-vad. vad_parameters: Dictionary of Silero VAD parameters or VadOptions class (see available parameters and default values in the class `VadOptions`). max_new_tokens: Maximum number of new tokens to generate per-chunk. If not set, the maximum will be set by the default max_length. chunk_length: The length of audio segments. If it is not None, it will overwrite the default chunk_length of the FeatureExtractor. clip_timestamps: Optionally provide list of dictionaries each containing "start" and "end" keys that specify the start and end of the voiced region within `chunk_length` boundary. vad_filter will be ignored if clip_timestamps is used. batch_size: the maximum number of parallel requests to model for decoding. hotwords: Hotwords/hint phrases to the model. Has no effect if prefix is not None. language_detection_threshold: If the maximum probability of the language tokens is higher than this value, the language is detected. language_detection_segments: Number of segments to consider for the language detection. Unused Arguments compression_ratio_threshold: If the gzip compression ratio is above this value, treat as failed. log_prob_threshold: If the average log probability over sampled tokens is below this value, treat as failed. no_speech_threshold: If the no_speech probability is higher than this value AND the average log probability over sampled tokens is below `log_prob_threshold`, consider the segment as silent. condition_on_previous_text: If True, the previous output of the model is provided as a prompt for the next window; disabling may make the text inconsistent across windows, but the model becomes less prone to getting stuck in a failure loop, such as repetition looping or timestamps going out of sync. Set as False prompt_reset_on_temperature: Resets prompt if temperature is above this value. Arg has effect only if condition_on_previous_text is True. Set at 0.5 prefix: Optional text to provide as a prefix at the beginning of each window. max_initial_timestamp: The initial timestamp cannot be later than this, set at 0.0. hallucination_silence_threshold: Optional[float] When word_timestamps is True, skip silent periods longer than this threshold (in seconds) when a possible hallucination is detected. set as None. Returns: A tuple with: - a generator over transcribed segments - an instance of TranscriptionInfo iThe current model is English-only but the multilingual parameter is set toTrue; setting to False instead.F sampling_rater!Processing audio with duration %sN)max_speech_duration_smin_silence_duration_msr)rrzPNo clip timestamps found. Set 'vad_filter' to True or provide 'clip_timestamps'.) max_durationTcPg|]"}fd|D#S)c<i|]\}}|t|zSr1)r>)rxkvrs r( zBBatchedInferencePipeline.transcribe...s,GGGtq!CM)**GGGr*)items)rxsegmentrs r(r|z7BatchedInferencePipeline.transcribe..sFHGGGw}}GGGr*rrzSSegment %d is longer than 30 seconds, only the first 30 seconds will be transcribed)rmr`segmentsc38K|]}|d|dz VdS)rrNr1)rxrs r( z6BatchedInferencePipeline.transcribe..s0SS''"22SSSSSSr*VAD filter removed %s of audiocXg|]&}j|dddf'S).Nr)rifeature_extractor)rxchunkr's r(r|z7BatchedInferencePipeline.transcribe..s6 U U UuTZ ) )% 0 0crc : U U Ur*enrgfloat32)dtypeaxisrrr,Detected language '%s' with probability %.2f`The current model is English-only but the language parameter is set to '%s'; using 'en' instead.rr^c,g|]}t|Sr1)r)rxfeatures r(r|z7BatchedInferencePipeline.transcribe..s CCCwk'**CCCr*rArBrCrDrErFrGrHrIrLrMrNrOrPrTrUrWrZrSrYrJrXrKrrVrQrRrhr^r_r`rarcrdrbr1)&rirris_multilingualloggerwarning isinstancenpndarrayrrinforrrrtkeyspopr RuntimeErrorrrrsumr concatenatefulln_melsr hf_tokenizerstackr@listtupleget_suppressed_tokensr]_batched_segments_generatorrestore_speech_timestamps)5r'rr^rrrArBrCrDrErFr<rIrGrHrJrKrMrNrOrPrQrRrSrTrUrVrrrWrrXrYrrZrrr`clip_timestamps_provided audio_chunksrrclip clip_durationrarrbr_rorrrrs5` @r(rz#BatchedInferencePipeline.transcribesv 4B  ! 0 @ ! J  % %2   !L%,, E mDDDE;q>M1  /1A(1K1K   $Ptz'C'P 7  !)%/.:03&&&NN 55..2E2E2G2GGG&**+BCCC%/&&(&&@L&&&N#8~"N"NL((-.u{1~#F#F"G"M (- $,:\--- )L// (, $.O -//L$_55  4##E$w-$u+*E$FGGG!%etG}!< M  2%%J%--H  &&"&w--"?$1%)F SS?SSS S S    , X(:: ; ;   "  U U U U U U U U  "  :#3 '($$ J..^ GTZ%5% # 21 # "6!5 #  21# !4 3# )D(C# kD%=99# BQB!]# *># 6!# "*>## (#%%iAAA$+# ."6!5/# 0!4 31# 2*>3# 4X5# 6,O7# 8-1D9# :(-u;# <,O=# >),?# @&A# B 21C# D#&#E# J!!51")&1   33         ( 0/=H~r*c#Ktt|| d}d}tdt||D]} ||| | |z||| | |z|} | D]} | D]} |dz }t | d|| dt | ddt | dd|jsdnd | d D| d | d | d | d|jd V|d| d|_ dS)Nr)totaldisablepositionrr5r6rrc&g|]}tdi|Sr1rrxr s r(r|zHBatchedInferencePipeline._batched_segments_generator..]s"!L!L!L4$,,,,!L!L!Lr*r;r7r8r:r9) r5r4r6rrr;r7r8r:r9r<rh) rrrrr3roundrSrLupdatecloserj) r'rrorrrrpbarseg_idxrrrrs r(rz4BatchedInferencePipeline._batched_segments_generatorDs#h--\1AANNNq#h--44  AllQ^+,A N 23 G"  %GqLG!$V_"$V_#GG$4a88!'%.!44$+#:MDD!L!L77;K!L!L!L&x0$+M$:'./?'@*12E*F$+$8$;$ A+ . %("""r*)r+r,r-rkrrrrr@r}r r0rr r[r>r.r r rrtrr3r]rrr1r*r(rfrfos^)))5!5!5!nN&*N&N&& N&N&N&N&f#' " !$%$%E E E 8;.2/2+/-0>B $#02t#''* %$2#E"<@(,&*04;?"&8;+,YDDS(BJ./D3-D D  D  DDDD"D"D5$u+uUCZ/@@AD(&.e_)D*%UO+D,&e_-D.%)/D0&+1D2!sHSM'9!:;3D4 5D67D8"$s),9D:!;D< %=D>?D@"ADB!CDDEDFGDH!tZ'7!89IDJ! KDLsmMDN"$t*-ODP*2%QDRSDT3-UDV'/uoWDX&)YDZ x "33 4[DDDDL %)%)%)%)%)r*rfcKeZdZ dadeded eeeefd ed ed ed eedede deedeeeeffdZ e deefdZ dbde fdZ dddddddddgddddddddddgdddd d!dddddd"ddddf"d#eeeejfd$eed%ed&ed'ed(ed)ed*ed+ed,ed-eeeeeed.ffd/eed0eed1eed2ed3ed4eeeeefd5eed6ed7eeed8ed9ed:ed;eded?eee efd@eedAeedBeeeefdCeedDeedEeedFedeeeeffHdGZdHedIeedJedKedLedMedeeefdNZ dbdOejdHedPedQeejdeef dRZdOejdejfdSZdQejdTeedHedPedeejj eeeff dUZ! dcdHedVeed8ed5eedDeedeef dWZ"dXee dHedQejdYed;eded?ee efdFedEedeeeeeeefffd`Z%dS)f WhisperModelautordefaultrNFmodel_size_or_pathdevice device_index compute_type cpu_threads num_workers download_rootlocal_files_onlyfilesrevisionuse_auth_tokenc t|_d\} }| r/|}| dd} | dd}n6tj|r|}nt |||| | }tjj |f|||||| d| |_ tj |d}| r%tj | |_nytj|r%tj ||_n5tj d|j jrdnd z|_||||_t-d i|j|_d |_|jj|jz|_|jj|jjz|_|jj|jz|_d |_d |_dS)a> Initializes the Whisper model. Args: model_size_or_path: Size of the model to use (tiny, tiny.en, base, base.en, small, small.en, distil-small.en, medium, medium.en, distil-medium.en, large-v1, large-v2, large-v3, large, distil-large-v2, distil-large-v3, large-v3-turbo, or turbo), a path to a converted model directory, or a CTranslate2-converted Whisper model ID from the HF Hub. When a size or a model ID is configured, the converted model is downloaded from the Hugging Face Hub. device: Device to use for computation ("cpu", "cuda", "auto"). device_index: Device ID to use. The model can also be loaded on multiple GPUs by passing a list of IDs (e.g. [0, 1, 2, 3]). In that case, multiple transcriptions can run in parallel when transcribe() is called from multiple Python threads (see also num_workers). compute_type: Type to use for computation. See https://opennmt.net/CTranslate2/quantization.html. cpu_threads: Number of threads to use when running on CPU (4 by default). A non zero value overrides the OMP_NUM_THREADS environment variable. num_workers: When transcribe() is called from multiple Python threads, having multiple workers enables true parallelism when running the model (concurrent calls to self.model.generate() will run in parallel). This can improve the global throughput at the cost of increased memory usage. download_root: Directory where the models should be saved. If not set, the models are saved in the standard Hugging Face cache directory. local_files_only: If True, avoid downloading the file and return the path to the local cached file if it exists. files: Load model files from the memory. This argument is a dictionary mapping file names to file contents as file-like or bytes objects. If this is set, model_path acts as an identifier for this model. revision: An optional Git revision id which can be a branch name, a tag, or a commit hash. use_auth_token: HuggingFace authentication token or True to use the token stored by the HuggingFace config folder. )NNztokenizer.jsonNpreprocessor_config.json)r" cache_dirr$r%)rrr intra_threads inter_threadsr#zopenai/whisper-tinyz.enr#g{Gz?ir1) rrrospathisdirr ctranslate2modelsWhisperrijoin tokenizersr from_bufferrisfile from_filefrom_pretrainedr_get_feature_kwargs feat_kwargsrr input_stride hop_lengthnum_samples_per_tokenrrwtokens_per_secondtime_precisionr)r'rrrrrr r!r"r#r$r% model_kwargstokenizer_bytespreprocessor_bytes model_pathtokenizer_files r(rkzWhisperModel.__init__msd!ll .8++  +J#ii(8$??O!&+Et!L!L   W]]- . . +JJ'"!1'!- J!'/   %%%%       j2BCC   * 4 @ @ Q QD   W^^N + +  * 4 > >~ N ND   * 4 D D%tz/I)TuU!!D  33J@RSS!1!E!ED4D!E!E  " -0A A "  " 0D4J4U U   " 0D4N N #r*rcH|jjrttndgS)z%The languages supported by the model.r)rirrrr&s r(supported_languagesz WhisperModel.supported_languagess#)- (BNtO$$$Nr*cJi} tj|d}|rtj|}n`tj|r?t |dd5}tj|}dddn #1swxYwYn|Sttj j fd| DS#tj$r%}|jd|Yd}~nd}~wwxYw|S)Nr'rutf-8)encodingc$i|] \}}|v || Sr1r1)rxrr valid_keyss r(rz4WhisperModel._get_feature_kwargs..s$GGGTQqJAqr*z&Could not load preprocessor config: %s)r,r-r2jsonloadsr5openloadrrrk parametersrrJSONDecodeErrorrr)r'rBrAconfig config_pathfileerKs @r(r8z WhisperModel._get_feature_kwargssg M',,z3MNNK! $677 ,, +sW===-!Yt__F--------------- "#3#<==HMMOOJGGGGV\\^^GGG G# M M M K   H! L L L L L L L L M sCA(C,-B C,BC,BC,AC,,D ;DD rrrrrrTrrrrr0rr^rrrArBrCrDrErFr<.rIrGrHrJrKrMrNrOrPrQrRrSrTrUrVrrrWrrXrYrZrrc$ ~4|jj4|r(|jjs|jdd}t |tjst|4}|j d4z }$|$}%|j dt|$|r |dkr|t}n!t |tr td2i|}t||}&t!||&\}'}(tj|'d}|j d4z }%|j d t|$|%z |jt&jr:|jd d 4fd |&Dnd}&||| })d}*d}+||jjsd}d},nt |t.r(t1|ddn|d}-|)j ddz }.|-|jz|.krt7|-|jznd}/||)d|/df|#|"\}},}+|j d||,n3|jjs%|dkr|jd|zd}d},t;|j|jj||}0t?d2id|d|d|d|d| d| d| d|d| d |d!|d"t | t@tBfr| n| gd#|d$|d%|d&|rtE|0|n|d'|d(|d)|d*|d+|d,|d-|d.|d/| d0|!}1|#|)|0|1||*}2|&rtI|2|&4}2tK||,|$|%|1||+1}3|2|3fS)3a'Transcribes an input file. Arguments: audio: Path to the input file (or a file-like object), or the audio waveform. language: The language spoken in the audio. It should be a language code such as "en" or "fr". If not set, the language will be detected in the first 30 seconds of audio. task: Task to execute (transcribe or translate). log_progress: whether to show progress bar or not. beam_size: Beam size to use for decoding. best_of: Number of candidates when sampling with non-zero temperature. patience: Beam search patience factor. length_penalty: Exponential length penalty constant. repetition_penalty: Penalty applied to the score of previously generated tokens (set > 1 to penalize). no_repeat_ngram_size: Prevent repetitions of ngrams with this size (set 0 to disable). temperature: Temperature for sampling. It can be a tuple of temperatures, which will be successively used upon failures according to either `compression_ratio_threshold` or `log_prob_threshold`. compression_ratio_threshold: If the gzip compression ratio is above this value, treat as failed. log_prob_threshold: If the average log probability over sampled tokens is below this value, treat as failed. no_speech_threshold: If the no_speech probability is higher than this value AND the average log probability over sampled tokens is below `log_prob_threshold`, consider the segment as silent. condition_on_previous_text: If True, the previous output of the model is provided as a prompt for the next window; disabling may make the text inconsistent across windows, but the model becomes less prone to getting stuck in a failure loop, such as repetition looping or timestamps going out of sync. prompt_reset_on_temperature: Resets prompt if temperature is above this value. Arg has effect only if condition_on_previous_text is True. initial_prompt: Optional text string or iterable of token ids to provide as a prompt for the first window. prefix: Optional text to provide as a prefix for the first window. suppress_blank: Suppress blank outputs at the beginning of the sampling. suppress_tokens: List of token IDs to suppress. -1 will suppress a default set of symbols as defined in `tokenizer.non_speech_tokens()`. without_timestamps: Only sample text tokens. max_initial_timestamp: The initial timestamp cannot be later than this. word_timestamps: Extract word-level timestamps using the cross-attention pattern and dynamic time warping, and include the timestamps for each word in each segment. prepend_punctuations: If word_timestamps is True, merge these punctuation symbols with the next word append_punctuations: If word_timestamps is True, merge these punctuation symbols with the previous word multilingual: Perform language detection on every segment. vad_filter: Enable the voice activity detection (VAD) to filter out parts of the audio without speech. This step is using the Silero VAD model https://github.com/snakers4/silero-vad. vad_parameters: Dictionary of Silero VAD parameters or VadOptions class (see available parameters and default values in the class `VadOptions`). max_new_tokens: Maximum number of new tokens to generate per-chunk. If not set, the maximum will be set by the default max_length. chunk_length: The length of audio segments. If it is not None, it will overwrite the default chunk_length of the FeatureExtractor. clip_timestamps: Comma-separated list start,end,start,end,... timestamps (in seconds) of clips to process. The last end timestamp defaults to the end of the file. vad_filter will be ignored if clip_timestamps is used. hallucination_silence_threshold: When word_timestamps is True, skip silent periods longer than this threshold (in seconds) when a possible hallucination is detected hotwords: Hotwords/hint phrases to provide the model with. Has no effect if prefix is not None. language_detection_threshold: If the maximum probability of the language tokens is higher than this value, the language is detected. language_detection_segments: Number of segments to consider for the language detection. Returns: A tuple with: - a generator over transcribed segments - an instance of TranscriptionInfo rFrrrrVNrrz0VAD filter kept the following audio segments: %sz, c3K|]9}dt|dz dt|dz dV:dS)[rz -> r]N)r)rxrrs r(rz*WhisperModel.transcribe..sq "E-U7^m-KLLLL,U5\M-IJJJJr*)rrr,r.rrrrrArBrCrDrErFrGrHrIrJrKrLrMrNrOrPrQrRrSrTrUrVrWrXrYrZrr1)&rrrirrrrrrrrrrrrtrrr isEnabledForloggingDEBUGdebugr2r0r.splitrwr>rrrr@rrrgenerate_segmentsrr])5r'rr^rrrArBrCrDrErFr<rIrGrHrJrKrMrNrOrPrQrRrSrTrUrVrrrWrrXrYrZrrr`ra speech_chunksrrrrrbr_start_timestampcontent_framesr5rorrrrs5 @r(rzWhisperModel.transcribesn.<  ! : ! K  2   !L%,, E mDDDE;q>M1%  /1A(1K1K     !/S00%!+ND11 >!+!=!=n!=!=1%HHM,:5-,P,P )L/N!>??    {'' 66  !!FII &3    !M))%l)KK!  :- '($$"/377,E///44Q7888(+ "*!3a!7')??.PP$*@@AAA((%c455j10K1M) (&   B( :- (d2B2B ##*,45 #$    J &    '! ! ! i! G! X! *> ! 21 ! "6!5 !  21! !4 3! )D(C! (B'A! )D(C!  *+e}EEX K=! *>! 6!! "*>#! (#%%iAAA$+! . 21/! 0#8"71! 2,O3! 4"6!55! 6!4 37! 8&9! :*>;! <,O=! >-L,K?! @XA! F)) i,    Y0=-XXH !51")&1   ~r*ror7rprqrrr5c g}tdko djcxko dknc}fdttD} t| dkrt| } |r"| td} | D]n} | | } | djz }| djz }|||jzz}|||jzz}|t |||| | } o|r||z }n| dz jz }|||jzz }n||}fdD}t|dkr+|djkr|djz }||jz}|t ||||z||z }|||fS) Nr#rcjg|]/}|dkr'|jkr|dz jk-|0S)rrtimestamp_begin)rxrror7s r(r|z>WhisperModel._split_segments_by_timestamps..sV" " " 1uuq Y666q1u !::: ;::r*r)r5rrr7rc*g|]}|jk |Sr1rhrxtokenros r(r|z>WhisperModel._split_segments_by_timestamps..;s-Ui6O-O-O-O-O-Or*)rrirrrr>rtr:)r'ror7rprqrrr5current_segmentsrconsecutive_timestampsslices last_slice current_slice sliced_tokensstart_timestamp_positionend_timestamp_position start_timeend_timelast_timestamp_positionr` timestampss `` r(rz*WhisperModel._split_segments_by_timestampss KK1  Ui.G!U!U!U!U6RT:!U!U!U!U " " " " " 3v;;''" " "  % & & * *011F& + c&kk***J!' + + &z-'? @ +8+;i>W+W()6r):Y=V)V&":T=P"PP')?$BU)UU ''!($, + & D $:>*Y-FF(/$2CCC(H#)J:""z"~9R'R'R*4R.9;T*T'2T5HH  # #%#h.!     L D'>>>r*rrrc#67K|jddz }t|jjz}t |jt r2d|jr|jdngD|_fd|jD}t|dkr| dt|dzdkr| |tt|ddd|ddd} d6d} d} | | d} g} d}|j {t |j t rGd |j z}||}| |n| |j t!|d | }d }| t| kr| | \}}||kr|}| |kr|} | |kr'| dz } | t| kr| | d} \| jjz}t| jjzjjz}t%jj|| z || z }|dd| | |zf}|jjz}t'|}jt,jr(jd t3|| |d}| dks||}|jr[j|}|dd\}}|dd}|j||_||_ !|||j"| dkr|j#nd|j$}%||||\} }!}"}#|j&S| j'|j&k}$|j( |!|j(krd}$|$r-jd| j'|j&| |z } a| j)d}%| }&dtTdtfd7dtVtTdtXf67fd }'dtZtTdtVtTfd}(.||%|||| \})} }*|j/rԉ0|)g||||j1|j2||*s.tg|)}+|+|+|krti|+j5z} |j6b|j6},|(|)}-|-8|'|-r-|-d|z }.|.|,kr|&ti|.j5zz} |}/tot|)D]}0|)|0}|ds|'|r|(|)|0dzd}1|1|1ddd}2n||z}2|d|/z |,kp|d|,kp|d|z dk}3|2|dz |,kp|'|1p||dz dk}4|3rJ|4rHtitq|dz|dj5z} ||dz |,kr|} g|)|0d<n |d}/tg|)}+|+|+}|)D]}|d}%|9|%}5|d|dks|5 sF| |%| dz } tu| |&|d|d|5|%|"|!|#| j'|j/rd |dDnd! V|j;r |"|j<kr7|j;r!jd"|"|j<t| }|=t%|| |&z jjz| t| k|>dS)#Nrrc,g|]}t|Sr1)r.)rxtss r(r|z2WhisperModel.generate_segments..[s.'''b '''r*r[c>g|]}t|jzSr1)rrw)rxr{r's r(r|z2WhisperModel.generate_segments..ds6" " " 35E"t-- . ." " " r*rr#u*"'“¿([{-"'.。,，!！?？:：”)]}、 seconds)r unitr rhzProcessing segment at %srf)rQrNrZFz$No speech threshold is met (%f > %f)r rc|dd}|d|dz }d}|dkr|dz }|dkr |d|z dzz }|d kr||d z z }|S) Nr!rhrrg333333?rg/$?@)get)r r!r`scores r(word_anomaly_scorez:WhisperModel.generate_segments..word_anomaly_scores{"hh}c:: ;g6%%SLEe##eh."44Ec>>X^+E r*rc||dsdSfd|dD}|dd}tfd|D}|dkp|dzt|kS)Nr;Fc(g|]}|dv |S)r r1)rxw punctuations r(r|zNWhisperModel.generate_segments..is_segment_anomaly..s'UUUq& 8T8T8T8T8Tr*rc3.K|]}|VdSNr1)rxrrs r(rzMWhisperModel.generate_segments..is_segment_anomaly..s/AAa..q11AAAAAAr*r g{Gz?)rr)rr;rrrs r(is_segment_anomalyz:WhisperModel.generate_segments..is_segment_anomalys?''*:? 5UUUUGG$4UUUbqb AAAA5AAAAAz?UT\SZZ%??r*rc6td|DdS)Nc3*K|]}|d |VdS)r;Nr1)rxss r(rzMWhisperModel.generate_segments..next_words_segment..s+??1AgJ?Q??????r*)next)rs r(next_words_segmentz:WhisperModel.generate_segments..next_words_segments!?????FFFr*rn)rjrr;rrr7c&g|]}tdi|Srrrs r(r|z2WhisperModel.generate_segments..Vs"CCC$CCCr*) r4r5rrr6r7r<r8r9r:r;zBReset prompt. prompt_reset_on_temperature threshold is met %f > %f)?rr.rtime_per_framerrXr0r`rrrr~rMstriprextendr nb_max_framesminrrr\r]r^r_rrVrirrorr^ language_coderrQrNrZgenerate_with_fallbackrHr:rGrrtr r[r rrSrrTrUrrrwrYrmaxrur3rJrKrr)8r'rrorrrrdcontent_duration seek_points seek_clipsidxclip_idxr5 all_tokensprompt_reset_sincerMinitial_prompt_tokensrrjseek_clip_start seek_clip_endrpwindow_end_timerqrrrrrrr_r^rrr8r<r9 should_skipr7 previous_seekrrrmr last_word_end threshold first_segmentgap hal_last_endsi next_segmenthal_next_startsilence_before silence_afterr6rrs8` @@r(razWhisperModel.generate_segmentsOs "+a/ $2H2W!WXX g-s 3 3 ''.G+11#666 '''G #" " " " 9@9P" " "  {  q   q ! ! ! {  a 1 $ $   ~ . . .,0  CCaC +add"3 4 4- - E (#A&   ! -'0#66 :!$w'='C'C'E'E!E(1(8(8(H(H%!!"78888!!'"8999* DTUUU # Z((-7-A *O]~-- . o%%&}$$A c*oo--%h/2D!7!FFK#.<<(78O&4%$L qqq$ )<"<<=G+d.D.SS !'**G{'' 66  !!.0@0M0M));)<)<=Oaxx>1!%W!5!5# 3*44^DD7>qz!}4 4)!B$/%.%8%D%D^%T%T "*2 '__#*#=)-w~~ ) %F++NFIwWW  !*6$3g6QQ .:#g&@@@#(K K%%>-3L(D)!,F M  %     @HTN @t @ @ @ @ @ @ @ GT$Z GHTN G G G G22#')!1 3   '&A :((%&" 0/*?)/M$+,<$=$=M$0][5P5P$]T5K%KLL:F ' GI%7$67G$H$HM$05G5G 5V5V0+G4{B??#05t?U9U3V3V#VD$$9L#C(8$9$9::66"22"6&w/%$--g66&+=+= 0a :,,L ,71=g1Fq1I'1R1 %f)zDLog probability threshold is not met with temperature %.1f (%f < %f)c|dS)Nrr1)xs r(z5WhisperModel.generate_with_fallback..s 1r*keyr ) r>rrRr>rWrrrrLrBrArCrirrDrErFrOrPrrrurrvrrIrr_rGrHr:r)r'rrror decode_result all_resultsbelow_cr_threshold_resultsrrr<kwargsrr7rrr8r6r9needs_fallbacks r(rz#WhisperModel.generate_with_fallbackzs  %'"&) '/$2EE F F' ' #  ! -Vw'==JJJ  ' 'Qs6{{QQF +QQ,6QQ7;oQQ ?Co QQQ #/` ` KQ!"&-o%&,7 ")!2 ' 0 )TZ(  '5#*#=%,%A%"&*&5 ' 7,G     F)!,F&kkG -*gw7M.MNK%15K##F++1133D 5d ; ; ! M   } - - -"N2>$w'JJJ%)NK%%`#); /55mDDD*6'"<<>-ABBM=!!T_%999 -.H10Dq0H.H I % 9 i7888 MM- ( ( ( r*r num_framesrjc  t|dkrdSg}g} |D]d} fd| D} |ttj| | | e||||} g} | D]\}tjd|D}|| }t|dkrtj |nd}tdt|}|dz}t|dkrd}tdt|D]{}||d ||d z |krX||d |vr||d |z||d <N||dz d |vr||d |z ||d <|t|||| ||f^t|D]x\}} d}| dd |jz }| |\}}t| D]>\}}d}g}|t| |kr|t| ||kr| ||}|d r`|t#|d t%||d zdt%||d zd|d |t|dz }|dz }|t| |kr|t| ||kt|dkr |dd |z |dzkr|dd |dd z |ks7t|dkr|dd |dd z |dzkrt|dkre|dd |dd z |krDt'|dd dz |dd |z }|x|dd <|dd <t'd|dd |z |dd <|d |dd krX|d dz |dd kr=t'dt|dd |z |d |dd <n|dd |d <|d |dd krJ|d dz|dd kr/t'|dd |z|d |dd <n|dd |d <|d }||||d<@z|S)Nrc8g|]}fd|dDS)c*g|]}|jk |Sr1)eotrks r(r|z?WhisperModel.add_word_timestamps...0s%RRR5EIM./sCSRRRJx$8RRRr*c0g|]}|d|dz S)rrr1rs r(r|z4WhisperModel.add_word_timestamps..<s%CCCetG},CCCr*rhgffffff?r#u .。!！?？rrrr r5r!)r rrr!r7rrr;)rrr itertoolschain from_iterablefind_alignmentrarraynonzeromedianrr.rmerge_punctuationsrrwrtrr)r'rrorrrTrUrj text_tokenstext_tokens_per_segmentrsegment_tokens alignmentsmedian_max_durations alignmentword_durationsmedian_durationrsentence_end_marksr segment_idx word_indexrpsubsegment_idxry saved_tokensr;timingboundarys ` r(rz WhisperModel.add_word_timestampsso x==A   F "$ ; ;G")N   tIO$A$A.$Q$QRR S S S # * *> : : : :(( {NJ   "# I IIXCCCCCN,N,B,B,D,DEN-0-@-@1-D-D .)))# "#u_'='=>>O*Q.L>""Q&&%3"q#i..11WWA |E*Yq\'-BB\QQ$Q</3EEE2;A,w2G,2VIaL//&q1u-f59KKK4=aL4G,4VIaL1 y*>@S T T T ' ',(G H H H H$-h$7$7J GJ G KJ!!*V,t/EEK,@,M )O\.7.@.@F GF G*   3z+'>#?#???LSV+K8HTTEE( 4Z@Ff~  %+F^&+K&/,I1&M&M$)+u *Eq$I$I,2=,A !Cx(8$9$99L!OJ!!3z+'>#?#???LSV+K8HTTEE(u::>>Qx-.0?!0CDDa%(7*;;lJJJJNN %a%(72C ClUVFV V V JJNN %a%(72C Cl R R'* %a! 3U1Xe_|5S((HCKJE!HUOeAhw.?,/58E?\3Q,R,Ra)#7+eAhuo==&w/#5a8III,/a/ A:gCVWW--a)) /4Ahw.? 7+#5)E"Ig,>>>&u-3eBi6FFF+.!"Ig.@*UBS,,b %((-2"Ie,< 5),6u,=)AF%n5g>>MF GN%$r*rmedian_filter_widthc 4t|dkrgS|j||j|||}g}t ||D]\}} |j|j} tjd| D} tjd| D} | | |j gz\} }t|dkr| gtj tj d|ddDd}t|dkr| gtj tj| dd t }| ||jz }||dd}||dd}fd t |dd|ddD}| d t | ||||D|S) Nr)rcg|] }|d Srr1rxpairs r(r|z/WhisperModel.find_alignment..$D$D$DT!W$D$D$Dr*cg|] }|d S)rr1rs r(r|z/WhisperModel.find_alignment..rr*rc,g|]}t|Sr1rrxts r(r|z/WhisperModel.find_alignment..s<<.sA"""Aq(1-.."""r*c Bg|]\}}}}}t|||||S))r r7rrr!)rt)rxr r7rrr!s r(r|z/WhisperModel.find_alignment..sR   >feS+!%#$/    r*)rrialignrr~rrrrsplit_to_word_tokensrrpadcumsumdiffastyper[r=)r'rorrrrr return_listr text_tokenr text_indices time_indicesr; word_tokensword_boundariesjumps jump_times start_times end_timesword_probabilitiesrs @r(rzWhisperModel.find_alignments {  q I*""   "   3 #   "%g{";";0 0  FJ%6 *J8$D$D$D$D$DEEL8$D$D$D$D$DEEL!*!?!?im_,"" E;;1$$ ""2&&& f <<;ss+;<<<==vO?##q((""2&&&F27<00&!LLLSSE&e,t/EEJ$_SbS%9:K"?122#67I"""" 4oabb6IJJ"""      BE{KDVBB       r*c | | Jd|g|r9t||}t||\}} tj|d}|d||jjz}||}|dd||jjzf}itd|jd|jjD]} | t|d| | |jjzf} |j | d} d| D} | d\}}||krnS |g|tfd }t|}||| fS) a Use Whisper to detect the language of the input audio or features. Arguments: audio: Input audio signal, must be a 1D float array sampled at 16khz. features: Input Mel spectrogram features, must be a float array with shape (n_mels, n_frames), if `audio` is provided, the features will be ignored. Either `audio` or `features` must be provided. vad_filter: Enable the voice activity detection (VAD) to filter out parts of the audio without speech. This step is using the Silero VAD model. vad_parameters: Dictionary of Silero VAD parameters or VadOptions class (see available parameters and default values in the class `VadOptions`). language_detection_threshold: If the maximum probability of the language tokens is higher than this value, the language is detected. language_detection_segments: Number of segments to consider for the language detection. Returns: language: Detected language. language_probability: Probability of the detected language. all_language_probs: List of tuples with all language names and probabilities. Nz.Either `audio` or `features` must be provided.rr.rc,g|]\}}|dd|fS)r#rfr1)rxrlprobs r(r|z0WhisperModel.detect_language.."s)!S!S!S-5$52;"5!S!S!Sr*c.t|Srr)langdetected_language_infos r(rz.WhisperModel.detect_language..-s%;D%A!B!Br*r)rrrrr n_samplesrrrrrrir setdefaultrr)r'rrrrrrrbrrrrrrbr^r_rs @r(rzWhisperModel.detect_languages>  !5!5 ;"6!5!5   = 5e^ L L 0>um0T0T- o|!<<<P-0F0PPPE--e44H U.1G1UUU U "$q(.,d.D.RSS I IA![[HS!a$2H2V.V*V%VWXXNj00@@CG"T!S7!S!S!S -?-B *H*#&BBB " - -h ; ; B BCW X X X X&BBBBH$''=h'G#H#H -/AAAr*) rrrrrNFNNNr)FNN)r)NNFNrr)&r+r,r-r0r r>r r r[rtrkpropertyrEr8rrrr.r rrr3r]rrrr@r/ StorageViewrarr0WhisperGenerationResultrrrrrr1r*r(rrls./%'+!&"&59eeeeCcN+ e  e  ee }eee3-e!sDy!12eeeeNOT#YOOOXO$*#' " !$%$%E E E 8;.2/2+/-0>B $#02t#('* %$2#E" <@(,&*36;?"&8;+,WSSS(BJ./S3-S S  S  SSSS"S"S5$u+uUCZ/@@AS(&.e_)S*%UO+S,&e_-S.%)/S0&+1S2!sHSM'9!:;3S4 5S67S8"$s),9S:!;S< %=S>?S@"ASB!CSDESFGSH!tZ'7!89ISJ! KSLsmMSNsDK/0OSP*2%QSR3-SST'/uoUSV&)WSX x "33 4YSSSSjM?M?S M? M?  M?  M?M? d3iM?M?M?M?j=A ^^*^^& ^ !!89 ^ ' ^^^^@ :rz :k.E : : : :@#/@S @ @ & @ {!95%N O @@@@L$) $"& !!!c!! !  ! 3- ! c!!!!FA%t*A%A%$/ A%  A% " A%!A% %A% A%A%A%A%R$% DDD#YD$/ D  D ! D dDDDDP'+)- 26+,.1IBIB #IB2:&IB IB dJ./ IB &) IB',IB sE4c5j 122 3IBIBIBIBIBIBr*rrrbrrc#6Kt||}|D]}|jrg}|jD]~}|j|jzdz }||}||j||_||j||_|||dj|_|dj|_||_n@||j|_||jd|_|VdS)Nr#rrT)is_end)rr;rrget_chunk_indexget_original_timer) rrbrts_maprr;r middle chunk_indexs r(rr4s! ! > >F = ME  # #*tx/14$44V<< #55dj+NN !33DHkJJ T""""!!HNGM)-GK!GMM#44W]CCGM 227;t2LLGK 'r*rcltj|}tj|}|Sr)rascontiguousarrayr/r from_array)rs r(rrQs-"7++G%0099G Nr*r6c|d}t|ttj|z S)NrH)rrzlibcompress)r6 text_bytess r(rvrvWs6W%%J z??Sz!:!:;; ;;r*rorPcd|vr'd|D}||jn7|t|dkrg}nt|ts Jd||j|j|j|j|j |j gttt|S)Nrcg|] }|dk| Srr1rs r(r|z)get_suppressed_tokens..as@@@a1r*rzsuppress_tokens must be a list)rnon_speech_tokensrrrr translatesotrsot_lm no_speechrsortedset)rorPs r(rr\s _@@o@@@y:;;;;  C$8$8A$=$=/400RR2RRRR    M            O,,-- . ..r*r prependedappendedct|dz }t|dz }|dkr||}||}|ddrO|d|vr3|d|dz|d<|d|dz|d<d|d<g|d<n|}|dz}|dkd}d}|t|kr||}||}|dds=|d|vr3|d|dz|d<|d|dz|d<d|d<g|d<n|}|dz }|t|kdSdS)Nr#rrr r}r7r+)r startswithrendswith)rr/r0rrprevious followings r(rrvs IA IA q&&Q<aL F  & &s + + 0@0F0F0H0HI0U0U ( 09V3D DIf "*8"4y7J"JIh !HV !#HX  A Q q&& A A c)nn  Q<aL ((-- )F2Cx2O2O'/)F2CCHV !)(!3i6I!IHX  "If "$Ih  A Q c)nn      r*)=rrLr]r,r# dataclassesrrinspectrmathrtypingrrr r r r warningsr r/numpyrr3rfaster_whisper.audiorr faster_whisper.feature_extractorrfaster_whisper.tokenizerrrfaster_whisper.utilsrrrrfaster_whisper.vadrrrrrr3r@r]rfrrtr>rrrrr0r.rvrrr1r*r(rAs  ))))))))CCCCCCCCCCCCCCCC::::::::======????????VVVVVVVVVVVV            ,  :  z)z)z)z)z)z)z)z)zEBEBEBEBEBEBEBEBP&w:g :RZK4K <<<<<< //3Z/d3i////4$t*PTr*