o
�J�h � @ s� d dl Z d dlZd dlmZmZmZmZmZ d dlZ d dl
Z
d dl Z d dl Z d dl
mZ d dlmZmZ d dlmZ d dlT G dd� d�ZdS ) � N)�List�Union�BinaryIO�Optional�Tuple)�DIARIZATION_MODELS_DIR)�DiarizationPipeline�assign_word_speakers)�
load_audio)�*c
@ s� e Zd Zefdefdd�Z ddeeeej f de
e dede e d e
e
e ef f
d
d �Z dde e de e fd d
�Zdd� Zedd� �Zedd� �ZdS )�Diarizer� model_dirc C s: | � � | _| �� | _d| _|| _tj| jdd� d | _d S )N�float16T��exist_ok) �
get_device�device�get_available_device�available_device� compute_typer
�os�makedirs�pipe)�selfr
� r �@C:\pinokio\api\whisper-webui.git\app\modules\diarize\diarizer.py�__init__ s
zDiarizer.__init__N�audio�transcribed_result�use_auth_tokenr �returnc
C s� t � � }|du r | j}|| jks| jdu r | j||d� t|�}| �|�}t|d|i�}g }|d D ]$} d}
d| v r>| d }
|
d | d �� } |�t| d | d | d
�� q2t � � | } || fS ) au
Diarize transcribed result as a post-processing
Parameters
----------
audio: Union[str, BinaryIO, np.ndarray]
Audio input. This can be file path or binary type.
transcribed_result: List[Segment]
transcribed result through whisper.
use_auth_token: str
Huggingface token with READ permission. This is only needed the first time you download the model.
You must manually go to the website https://huggingface.co/pyannote/speaker-diarization-3.1 and agree to their TOS to download the model.
device: Optional[str]
Device for diarization.
Returns
----------
segments_result: List[Segment]
list of Segment that includes start, end timestamps and transcribed text
elapsed_time: float
elapsed time for running
N)r r �segments�None�speaker�|�text�start�end)r& r' r% ) �timer r � update_piper
r �strip�append�Segment)
r r r r r �
start_timeZdiarization_segmentsZdiarized_result�segments_result�segmentr# Z
diarized_text� elapsed_timer r r �run s6 �
�
� z Diarizer.runc C sn |du r| � � }|| _tj| jdd� t�| j�s!|s!td� dS t�d�}d|_ t
||| jd�| _ d|_ dS )a�
Set pipeline for diarization
Parameters
----------
use_auth_token: str
Huggingface token with READ permission. This is only needed the first time you download the model.
You must manually go to the website https://huggingface.co/pyannote/speaker-diarization-3.1 and agree to their TOS to download the model.
device: str
Device for diarization.
NTr z�
Failed to diarize. You need huggingface token and agree to their requirements to download the diarization model.
Go to "https://huggingface.co/pyannote/speaker-diarization-3.1" and follow their instructions to download the model.
z speechbrain.utils.train_logger)r r � cache_dirF) r r r r r
�listdir�print�logging� getLogger�disabledr r )r r r �loggerr r r r) X s& ��
�
zDiarizer.update_pipec C sf | j dur
| ` d| _ | jdkrtj�� tj�� | jdkr-tj�� tj�� tj�� t �
� dS )z(Offload the model and free up the memoryN�cuda�xpu) r r �torchr9 � empty_cache�reset_max_memory_allocatedr: � reset_accumulated_memory_stats�reset_peak_memory_stats�gc�collect)r r r r �offload s
zDiarizer.offloadc C s0 t j�� rdS t j�� rdS t jj�� rdS dS )Nr9 r: �mps�cpu)r; r9 � is_availabler: �backendsrC r r r r r � s
zDiarizer.get_devicec C sH dg} t j�� r
| �d� t j�� r| �d� t jj�� r"| �d� | S )NrD r9 r: rC )r; r9 rE r+ r: rF rC )�devicesr r r r � s
z Diarizer.get_available_device)N)NN)�__name__�
__module__� __qualname__r �strr r r �np�ndarrayr r, r r �floatr1 r) rB � staticmethodr r r r r r r s: �
������
�?��
�'
r )r r; �typingr r r r r �numpyrL r( r5 r@ �modules.utils.pathsr Z modules.diarize.diarize_pipeliner r � modules.diarize.audio_loaderr
� modules.whisper.data_classesr r r r r �<module> s