Speaker ID + Diarization FT: pyannote.audio + WavLM — Çoklu Konuşmacı Ayrımı

Toplantı/çağrı merkezi transkripti: 'kim konuşuyor + ne diyor'. pyannote.audio (HF), WavLM speaker embedding, diarization pipeline (VAD → embedding → clustering). Çağrı merkezi case: müşteri vs operatör ayrımı, RTX 4090 + 100 saat TR çağrı dataset üzerinde FT.

Şükrü Yusuf KAYA

24 dakikalık okuma

14.05.2026

İleri

Speaker ID + Diarization FT: pyannote.audio + WavLM — Çoklu Konuşmacı Ayrımı

python

# === pyannote.audio + Whisper diarization pipeline ===
from pyannote.audio import Pipeline
import whisper
 
# Diarization
diar = Pipeline.from_pretrained("pyannote/speaker-diarization-3.1",
                                  use_auth_token="hf_xxx")
 
# ASR
asr = whisper.load_model("large-v3-turbo")
 
def transcribe_with_speakers(audio_path):
    # 1. Diarization
    diarization = diar(audio_path)
 
    # 2. Per-segment ASR
    result = []
    for turn, _, speaker in diarization.itertracks(yield_label=True):
        # Extract audio slice
        # ... segment WAV file generate ...
        text = asr.transcribe(segment_audio, language="tr")["text"]
        result.append({
            "speaker": speaker,
            "start": turn.start,
            "end": turn.end,
            "text": text,
        })
    return result
 
# Output:
# [{"speaker": "SPEAKER_00", "start": 0.0, "end": 3.4, "text": "Merhaba, nasıl yardımcı olabilirim?"},
#  {"speaker": "SPEAKER_01", "start": 3.4, "end": 8.2, "text": "Telefonumda problem var."},
#  ...]