o 3NPit5@sJdZddlZddlZddlmZddlmZmZmZm Z m Z ddl Z ddl Z ddlZddlmZedZd!deed ed ed efd d Z  d"dedeeded eee eefeeffddZddZ   d#de eefdedededed eeeeeff ddZ d$deeeeefded eeeeeffdd ZdS)%a Speaker diarization with GPU optimization and compute monitoring. === OPTIMIZATION (v6.2) === Key improvements: - In-memory chunk processing: Pass waveform tensors directly to pyannote - No disk writes: Eliminates chunk WAV file I/O - VAD-aware chunking (cut at silence) - Compute utilization tracking - Memory-efficient processing === EDGE CASE HANDLING (v6.7) === - Retry logic for transient GPU failures - OOM recovery with cache clearing - Graceful fallback on persistent failures N)Path)ListDictTupleOptionalUnion)MODELSzFastPipelineV6.Diarization>@ vad_segments target_timewindowreturnc Csr|}d}tt|dD]*}||d}||dd}||}||d} t| ||kr6||kr6|}| }q |S)u Find silence point near target time for chunk boundary. Strategy: Find largest gap within ±window of target_time. This ensures we don't cut in the middle of speech. endstart)rangelenabs) r r r best_cutbest_gapi gap_startgap_endgap_size gap_centerrU/home/ubuntu/.cursor/worktrees/maya3data__SSH__216.81.248.184_/zxg/src/diarization.pyfind_vad_safe_cuts  rT audio_path in_memorycCstd|rdndd|dur |j}|j}t|d}nt|\}}| d }|j d|}|sOt |j d} | jd d dd lm} | | g} d } d} | |krt| |j|}||krzt||}|| |jkryt| |j|}n|}t| |}t||}|dd||f}|r||d }| || || fn| d| dd}tt|||| t|| || f|} | d7} | |ksY|rdnd}tdt| d|d| S)an Create VAD-aware audio chunks for diarization. === OPTIMIZATION (v6.2) === - in_memory=True (default): Return waveform tensors directly (no disk I/O!) - in_memory=False: Legacy disk-based WAV files Strategy: 1. Target chunk size from config 2. Find optimal cut points at silence boundaries 3. Never cut in the middle of speech Args: audio_path: Path to audio file vad_segments: Speech segments from VAD config: Pipeline configuration audio_buffer: Optional AudioBuffer (avoids file re-read) in_memory: If True, return dicts with waveform tensors instead of file paths Returns: List of (chunk_data, start_time, duration) where chunk_data is either: - str: file path (in_memory=False) - dict: {"waveform": tensor, "sample_rate": int} (in_memory=True) u"✂️ Creating VAD-aware chunks (z in-memorydisk)...NrrchunksT)exist_ok) TEMP_MANAGERr)waveform sample_ratechunk_03dz.wavzto disku ✅ Created z chunks (z, VAD-aware boundaries))loggerinfo waveform_npr(torch from_numpy unsqueeze torchaudioloadsqueezenumpyshaperparentmkdirsrc.audio_bufferr&registerminchunk_durationrmin_chunk_durationintappendsavestrr)r r config audio_bufferr!r-srr'total_duration chunk_dirr&r$ridx target_endr start_sample end_samplechunk_waveform chunk_data chunk_pathmode_strrrr create_chunks6sR       "rNcCst|dr|jS|S)z Extract Annotation object from diarization result. === PYANNOTE 4.x COMPATIBILITY === - pyannote 3.x: Returns Annotation directly - pyannote 4.x: Returns DiarizeOutput with .speaker_diarization attribute speaker_diarization)hasattrrO)resultrrr_get_annotation_from_results rRFrrK chunk_startextract_overlaps chunk_idx max_retriesc Csd}t|dD]S}ztj||j|jd}t|} t|tr%t|j } nd|d} g} |rddl m } | | |\} }| D]}| ||d||d |d d d q>g}|D](}||d}||d }||}||j kr| |||| d |ddqYn0g}| jddD]'\}}}||j}||j}||}||j kr| |||| d |dq|| fWSty'}za|}t|}d|vsd|vrtd|d|ddtjddtjrtjtj||krtdWYd}~qntd|d|||krWYd}~qWYd}~qd}~wty\}z)|}td|d|dd|||krRtjddWYd}~qWYd}~qd}~wwtd|d|dd|ggfS)a Diarize a single chunk using the hot model. === OPTIMIZATION (v6.2) === Supports both file paths AND in-memory waveform dicts. No more disk I/O for chunks! === PYANNOTE 4.x COMPATIBILITY === Handles both pyannote 3.x (returns Annotation) and 4.x (returns DiarizeOutput). === EDGE CASE HANDLING (v6.7) === - Retry logic for transient GPU failures - OOM recovery with cache clearing - Graceful fallback on persistent failures Uses global MODELS singleton to avoid reloading. Returns segments with global timestamps, and optionally overlap regions. Args: chunk_data: Either file path (str) OR dict {"waveform": tensor, "sample_rate": int} chunk_start: Start time offset for global timestamps config: Pipeline configuration extract_overlaps: If True, also extract overlap regions (for unified OSD+diarization) chunk_idx: Chunk index for speaker naming (used for in-memory chunks) max_retries: Number of retries on failure (default: 2) Returns: (segments, overlaps): segments list, and overlaps list (empty if extract_overlaps=False) Nr) min_speakers max_speakersr)r*r)!extract_overlaps_from_diarizationrrdurationoverlap)rrrZtype_speaker)rrrZr^T) yield_labelz out of memorycuda Chunk z attempt z(: GPU error, clearing cache and retrying aggressivez RuntimeError: z failed: zChunk z diarization failed after z attempts: )rrdiarization_pipelinerWrXrR isinstancer@rstemsrc.overlap_detectionrYr>min_segment_duration itertracksrr RuntimeErrorlowerr+warning clear_cacher.r` is_available empty_cache synchronizetimesleeperror Exception)rKrSrArTrUrV last_errorattempt raw_resultrQchunk_idoverlapsrYchunk_overlapsclean_segmentsovlsegmentsseg global_start global_endrZturnr]r^e error_strrrr diarize_chunks%                              rr$cCs|rdnd}td|dt|dtjdt}g}g}t|D]a\}\}} } t} t|| |||d\} } || || t| }|rTdt| d nd}td |d d t|d t| d|d|dd| dd |d |j dkrtj ddq$tj dd|j ddd|j dddt|}tt dd|D}|rdt|d nd}td|d |ddt|d|d|d ||fS)!ae Run diarization on all chunks with progress tracking. === OPTIMIZATION (v6.1) === When extract_overlaps=True, this function extracts overlap regions from the diarization results in a SINGLE PASS, eliminating the need for a separate OSD stage that would run diarization again. Args: chunks: List of (chunk_path, chunk_start, chunk_duration) config: Pipeline configuration extract_overlaps: If True, also extract and return overlap regions Returns: (all_segments, all_overlaps): segments and overlaps (overlaps empty if not extracted) z + OSDu🎙️ Running diarizationz (z chunks, model: r#)rTrUz, z overlapsrar/z: z segmentsz.1fzs, z.0fzs audio)rTrbcS|dSNrrxrrrbz!run_diarization..)keycSrrrrrrrrcrcss|]}|dVqdS)r^Nr).0srrr fsz"run_diarization..z | u✅ Diarizationzs | z | ~z fragments) r+r,rr _model_namerq enumeraterextendclear_cache_every_n_chunksrmsortset)r$rArTrMr all_segments all_overlapsrrKrS chunk_durt0r}ryelapsed overlap_strunique_speakersrrrrun_diarization.s8 $   @   0r)r )NT)Frr)F)__doc__rqloggingpathlibrtypingrrrrrr4npr.r1 src.modelsr getLoggerr+floatrr@boolrNrRr=rrrrrrs`     _