AbijahKaj / whisper-telephony-amd

Whisper Telephony AMD (Answering Machine Detection)

Streaming audio classifier that detects whether you're talking to a human , voicemail , IVR , or answering machine — using Whisper's speech understanding to distinguish human-recorded voicemail from live speech.

Quick Start

Install

pip install transformers datasets evaluate torch torchaudio soundfile accelerate
huggingface-cli login  # needed for private dataset

Train on your 5090

python train_local.py
# Or customize:
python train_local.py --model openai/whisper-medium --epochs 30 --batch-size 16

Streaming Inference

from streaming_amd import StreamingAMDClassifier

classifier = StreamingAMDClassifier("AbijahKaj/whisper-telephony-amd")

for pcm_chunk in audio_stream:  # 160ms chunks @ 8kHz
    result = classifier.process_chunk(pcm_chunk)
    if result:
        label, confidence, elapsed_s = result
        print(f"{label} ({confidence:.0%}) after {elapsed_s:.1f}s")
        break

Architecture

• Base : openai/whisper-small encoder (244M params)
• Why Whisper : Voicemail greetings are recorded by humans — sound identical to live speech. Whisper understands what is being said ("leave a message" vs "hello?")
• Streaming : Accumulating PCM buffer → inference every 500ms → confidence-gated output
• Classes : human (0), voicemail (1), ivr (2), answering_machine (3)

Dataset

AbijahKaj/telephony-amd-dataset (private):

• 1,599 samples (~3.5 hours), balanced across 4 classes
• 200 real human voices from PolyAI/MINDS14 (telephony callers)
• 1,399 TTS-generated via Microsoft Neural TTS (edge-tts)
• Telephony channel effects: 300-3400Hz bandpass, noise, gain variation
• Voicemail/AM samples include beep tones; IVR includes DTMF

Files

• train_local.py — Training script with CLI args (5090-ready)
• streaming_amd.py — Streaming inference module
• build_dataset.py — Dataset generation script (edge-tts + MINDS14)