^jA d Z ddlZddlmZ ddlmZmZmZ ddl m
Z
ddl m Z ddl
mZmZ e rddlZ ej" e Z G d d
e
Zd
gZy) z%
Feature extractor class for Whisper
N )is_torch_available)mel_filter_bank spectrogramwindow_function)SequenceFeatureExtractor) BatchFeature)
TensorTypeloggingc e Zd ZdZdgZ d fd Zdej dedej fdZ ddej dedej fd Z
e dd
e ej d e ej d e
de ej fd
Z d dej e e
z e ej z e e e
z dededz deez dz dedz dedz dedz dedz dedz dedz defdZ xZS ) WhisperFeatureExtractora
Constructs a Whisper feature extractor.
This feature extractor inherits from [`~feature_extraction_sequence_utils.SequenceFeatureExtractor`] which contains
most of the main methods. Users should refer to this superclass for more information regarding those methods.
This class extracts mel-filter bank features from raw speech using a custom numpy implementation of the `Short Time
Fourier Transform` which should match pytorch's `torch.stft` equivalent.
Args:
feature_size (`int`, *optional*, defaults to 80):
The feature dimension of the extracted features.
sampling_rate (`int`, *optional*, defaults to 16000):
The sampling rate at which the audio files should be digitalized expressed in hertz (Hz).
hop_length (`int`, *optional*, defaults to 160):
Length of the overlapping windows for the STFT used to obtain the Mel Frequency coefficients.
chunk_length (`int`, *optional*, defaults to 30):
The maximum number of chunks of `sampling_rate` samples used to trim and pad longer or shorter audio
sequences.
n_fft (`int`, *optional*, defaults to 400):
Size of the Fourier transform.
padding_value (`float`, *optional*, defaults to 0.0):
Padding value used to pad the audio. Should correspond to silences.
dither (`float`, *optional*, defaults to 0.0):
Adds dithering. In other words, adds a small Gaussian noise to each frame.
E.g. use 0.0001 to add dithering with a normal distribution centered
around 0.0 with standard deviation 0.0001 (assuming [-1,+1] range of raw_speech).
The value 0.0 means no dithering.
Dithering has similar effect as `spectrogram(mel_floor=...)`. It reduces
the high log_mel_fbank values for signals with hard-zero sections,
when VAD cutoff is present in the signal.
input_featuresc t
| d||||d| || _ || _ || _ ||z | _ | j
|z | _ || _ || _ t d|dz z |dd|dd | _
y ) N) feature_size
sampling_rate
padding_valuereturn_attention_mask g @@slaney)num_frequency_binsnum_mel_filters
min_frequency
max_frequencyr norm mel_scale ) super__init__n_fft
hop_length chunk_length n_samples
nb_max_framesr ditherr mel_filters) selfr r r" r# r! r r& r kwargs __class__s /var/www/ramen.bs-engineer-server.com/venv/lib/python3.12/site-packages/transformers/models/whisper/feature_extraction_whisper.pyr z WhisperFeatureExtractor.__init__E s
%''"7
$ (%
5 !^^z 9 * * 5A:~ ( '
waveform_batchdevicereturnc |dk7 rt d| d g }|D ] }t |t | j d | j | j d| j
| j d }|dddd f }t j ||j d
z
}|d z d z }|j | t j | }|S ) z
Compute the log-mel spectrogram of the provided audio, gives similar results to Whisper's original torch
implementation with 1e-5 tolerance.
cpuz Got device `z` for feature extraction, but feature extraction on CUDA accelerator devices requires torch, which is not installed. Either set `device='cpu'`, or install torch according to the official instructions: https://pytorch.org/get-started/locally/hanng @log10) frame_lengthr" powerr& r' log_melN @ @)
ValueErrorr r r! r" r& r' npmaximummaxappendarray)r( r- r. log_spec_batchwaveformlog_specs r+ _np_extract_fbank_featuresz2WhisperFeatureExtractor._np_extract_fbank_featuresi s
U? vh 'q q
& ,H"
F3 !ZZ ??{{ , , H 3B3'Hzz(HLLNS,@AH 3#-H ! !( + , .1 r, rA c 8 t j | j |t j }t j | j
| }| j dk7 rF|| j t j |j |j |j z z
}t j || j
| j |d }|dddf j d z }t j | j j |t j }|j |z }t j |d
j# }|j% d k( rD|j' d d d
j' dd d
} t j( || dz
}n't j( ||j' dz
}|dz dz }|dk7 r |j+ j- }|j/ S )z
Compute the log-mel spectrogram of the audio using PyTorch's GPU-accelerated STFT implementation with batching,
yielding results similar to cpu computing with 1e-5 tolerance.
)r. r )dtyper. T)windowreturn_complex.Nr7 r g|=)min)dimkeepdimr r r8 r9 r1 )torch
from_numpytofloat32 hann_windowr! r& randnshaperE r. stftr" absr' Tclampr3 rI r= r< detachr1 numpy)
r( rA r. rF rR
magnitudesr' mel_specrB max_vals
r+ _torch_extract_fbank_featuresz5WhisperFeatureExtractor._torch_extract_fbank_features s
##H-00 G""4::f=
;;# ekk(.._g_n_n&oo oHzz(DJJ_cd#ss(^'')Q.
&&t'7'78;;FEMMR ==:-;;xU399; <<>Q llq$l7:>>1d>STUVG }}Xw}=H }}Xx||~/CDH sNc) U? (,,.H~~ r, input_valuesattention_maskr c |t j |t j }g }t | |j d D ]m \ }}||d| j z
t j |d| j dz z }||j d k r|||d |j | o |S | D cg c]< }||j z
t j |j dz z > }}|S c c}w )z[
Every array in the list is normalized to have zero mean and unit variance
Nr7 gHz>r )
r; r? int32zipsummeansqrtvarrQ r> )r\ r] r normed_input_valuesvectorlength normed_slicexs r+ zero_mean_unit_var_normz/WhisperFeatureExtractor.zero_mean_unit_var_norm s % XXnbhh ?N"$ "%lN4F4Fr4J"K
9 & )=)=)? ?2776RYSY?K^K^K`cgKgCh h L . .q 11,9L)#**<8
9 #" Vb"bPQALBGGAEEGdN4K#K"b "b"" #cs :AC?N
max_length
raw_speech
truncationpad_to_multiple_ofreturn_tensorsr paddingr do_normalizec f |O|| j k7 rmt d| j j d| j d| j d| d t j d| j j d t
|t j xr t |j d kD } | r&t |j d
kD rt d | | xs@ t
|t t f xr( t
|d t j t t f }
|
r>|D cg c]2 }t j |gt j
j 4 }}n|
s@t
|t j s&t j |t j
}nht
|t j rN|j t j t j" u r |j% t j }|
s!t j |g j g}t' d|i }| j) |||r|n | j* |||xs | }| rD| j- |d |d | j. |d< t j0 |d d |d< |j3 d j5 d
d d }t7 r | j8 n | j: } ||d |
}t
|d t r7|D cg c]' }t j |t j
) c}|d< n||d< |rK|d dddd| j< f }|d j d | j< z d k7 r |ddddf }||d< ||j? | }|S c c}w c c}w )a] Main method to featurize and prepare for the model one or several sequence(s). Implementation uses PyTorch
for the STFT computation if available, otherwise a slower NumPy based one.
Args:
raw_speech (`np.ndarray`, `list[float]`, `list[np.ndarray]`, `list[list[float]]`):
The sequence or batch of sequences to be padded. Each sequence can be a numpy array, a list of float
values, a list of numpy arrays or a list of list of float values. Must be mono channel audio, not
stereo, i.e. single float per timestep.
truncation (`bool`, *optional*, default to `True`):
Activates truncation to cut input sequences longer than *max_length* to *max_length*.
pad_to_multiple_of (`int`, *optional*, defaults to None):
If set will pad the sequence to a multiple of the provided value.
This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability
`>= 7.5` (Volta), or on TPUs which benefit from having sequence lengths be a multiple of 128.
return_tensors (`str` or [`~utils.TensorType`], *optional*):
If set, will return tensors instead of list of python integers. Acceptable values are:
- `'pt'`: Return PyTorch `torch.Tensor` objects.
- `'np'`: Return Numpy `np.ndarray` objects.
return_attention_mask (`bool`, *optional*):
Whether to return the attention mask. If left to the default, will return the attention mask according
to the specific feature_extractor's default.
[What are attention masks?](../glossary#attention-mask)
For Whisper models, `attention_mask` should always be passed for batched inference, to avoid subtle
bugs.
padding (`str` or [`~utils.PaddingStrategy`], *optional*, defaults to `'max_length'`):
Activates and controls padding. Accepts the following values:
- `'longest'`: Pad to the longest sequence in the batch (or no padding if only a single sequence is
provided).
- `'max_length'` (default): Pad to a maximum length specified with the argument `max_length` or to the
maximum acceptable input length for the model if that argument is not provided.
- `'do_not_pad'`: No padding (i.e., can output a batch with sequences of different lengths).
max_length (`int`, *optional*):
Controls the maximum length to use by one of the truncation/padding parameters.
If left unset or set to `None`, this will use the predefined model maximum length if a maximum length
is required by one of the truncation/padding parameters. If the model has no specific maximum input
length (like XLNet) truncation/padding to a maximum length will be deactivated.
sampling_rate (`int`, *optional*):
The sampling rate at which the `raw_speech` input was sampled. It is strongly recommended to pass
`sampling_rate` at the forward call to prevent silent errors and allow automatic speech recognition
pipeline.
do_normalize (`bool`, *optional*, defaults to `False`):
Whether or not to zero-mean unit-variance normalize the input. Normalizing can help to significantly
improve the performance of the model.
device (`str`, *optional*, defaults to `'cpu'`):
Specifies the device for computation of the log-mel spectrogram of audio signals in the
`_torch_extract_fbank_features` method. (e.g., "cpu", "cuda")
**kwargs: Not supported by WhisperFeatureExtractor.__call__() and ignored.
Nz3The model corresponding to this feature extractor: z& was trained using a sampling rate of zI. Please make sure that the provided `raw_speech` input was sampled with z and not .zDIt is strongly recommended to pass the `sampling_rate` argument to `zN()`. Failing to do so can result in silent errors that might be hard to debug.r r z2Only mono-channel audio is supported for input to r )rE r )rp rk rm rn r r] )r] r )axisr7 ) r r: r* __name__loggerwarning
isinstancer; ndarraylenrQ listtupleasarrayrN rT rE float64astyper padr$ rj r stackget transposer r[ rC r" convert_to_tensors)r( rl rm rn ro r rp rk r rq r. r) is_batched_numpy
is_batchedspeechbatched_speech
padded_inputsr extract_fbank_featuresfeaturerescaled_attention_masks r+ __call__z WhisperFeatureExtractor.__call__ s P $ 2 22 I$..JaJaIb c))-););(< =))-););( " 0 0 /K /M* +
/1hh}EU7V]^._M* + '**+;<FFq!QO 3E2FD . .DLkLk 0q0A6J nQ' .dr.sY`rzz'/T.sM* + /=M* + &34D&EaI[DOOI[F[&\ #
-.44Q7$//IQN*A!SbS&*I'.EM* + % ) < <^ LM y ]R /ts %7N),N.)P i> i r r F)r1 )r ) TNNNrk NNNr1 )ru
__module__ __qualname____doc__model_input_namesr r; ry strrC r[ staticmethodr{ floatrj boolintr
r r
__classcell__)r* s @r+ r
r
! s B ** #"
H S UWU_U_ < bjj # Z\ZdZd > be#2::r sX " H H I 4 ( H %u 6 u p %
%r,