o i @svddlZddlmZddlmZddlmZddlZddl mZ ddl Z ddl m Z zddlZWn ey=e dZYnwzddlmZWneyVe ddZYnwGd d d eeZeGd d d Zed ejdZeddZde jeje jBdede jeje jBfddZde jejdedede jejfddZde jejdedefddZ GdddZ!dS)N) dataclass)Enum)Literal)PlaceholderModulelibrosascipysignalc@s eZdZdZdZdZdZdZdS)ChannelReductionz8Method to reduce multi-channel audio to target channels.meanfirstmaxsumN)__name__ __module__ __qualname____doc__MEANFIRSTMAXSUMrrK/home/ubuntu/vllm_env/lib/python3.10/site-packages/vllm/multimodal/audio.pyr s r c@sPeZdZUdZdZedBed<ejZ eed<e de fddZ de fd d ZdS) AudioSpecaSpecification for target audio format. This dataclass defines the expected audio format for a model's feature extractor. It is used to normalize audio data before processing. Attributes: target_channels: Number of output channels. None means passthrough (no normalization). 1 = mono, 2 = stereo, etc. channel_reduction: Method to reduce channels when input has more channels than target. Only used when reducing channels. Ntarget_channelschannel_reductionreturncCs |jduS)z&Whether audio normalization is needed.Nrselfrrrneeds_normalization6s zAudioSpec.needs_normalizationcCs&|jdurdSd|jd|jjdS)NzAudioSpec(passthrough)zAudioSpec(channels=z , reduction=))rrvaluerrrr__repr__;s  zAudioSpec.__repr__)rrrrrint__annotations__r rrpropertyboolr strr#rrrrr%s  rr)rrraudiospecrcCs|js|S|jdkr|jdkr|Std|jd|jdkr(td|jd|jd|jdkr>t|tjr;|jn|j}|jd}||jkrJ|S||jkrZtd|d |jt|tj}|jdkr|j t j kr~|rvtj |dd }|S|j dd }|S|j t j kr|d}|S|j t jkr|rtj|dd }|S|jdd j}|S|j t jkr|rtj|dd }|S|jdd }|Std |j |d |jS)aKNormalize audio to the specified format. This function handles channel reduction for multi-channel audio, supporting both numpy arrays and torch tensors. Args: audio: Input audio data. Can be: - 1D array/tensor: (time,) - already mono - 2D array/tensor: (channels, time) - standard format from torchaudio - 2D array/tensor: (time, channels) - format from soundfile (will be auto-detected and transposed if time > channels) spec: AudioSpec defining the target format. Returns: Normalized audio in the same type as input (numpy or torch). For mono output (target_channels=1), returns 1D array/tensor. Raises: ValueError: If audio has unsupported dimensions or channel expansion is requested (e.g., mono to stereo). rzCannot expand mono audio to z channelszUnsupported audio shape: z. Expected 1D or 2D.rzCannot expand z channels to )axis)dimzUnknown reduction method: N)r ndimr ValueErrorshape isinstancenpndarrayTrr rr rrr valuesrr )r)r* num_channelsis_numpyresultrrrnormalize_audioIsN              r9orig_sr target_srcCstj|||dS)Nr:r;)rresampler)r:r;rrrresample_audio_librosasr?cCs8||kr t|d||S||krt|||dS|S)Nr) scipy_signal resample_polyr>rrrresample_audio_scipys rBc@sVeZdZdZ  ddedBdedfddZd eje j d ed eje j fd d Z dS)AudioResamplerz,Resample audio data to a target sample rate.Nrr;method)rrcCs||_||_dS)N)r;rD)rr;rDrrr__init__s zAudioResampler.__init__r)r:rcCsx|jdur tdtjt|t|jdddr|S|jdkr&t|||jdS|jdkr3t|||jdStd|jd ) NzBAudio resampling is not supported when `target_sr` is not providedggư>)rel_tolabs_tolrr<rzInvalid resampling method: z.. Supported methods are 'librosa' and 'scipy'.) r; RuntimeErrormathisclosefloatrDr?rBr/)rr)r:rrrr=s,    zAudioResampler.resample)Nr) rrrrrKrrEnptNDArrayr2floatingr=rrrrrCs    rC)"rI dataclassesrenumrtypingrnumpyr2 numpy.typingrLtorchvllm.utils.import_utilsrr ImportError scipy.signalrr@placeholder_attrr(r rrMONO_AUDIO_SPECPASSTHROUGH_AUDIO_SPECrMrNTensorr9rKr?rBrCrrrrs^            S