AbijahKaj/telephony-amd-dataset

Telephony AMD (Answering Machine Detection) Dataset

Overview

A multilingual 4-class telephony audio classification dataset for training streaming Answering Machine Detection models. Contains real human speech (PolyAI/MINDS14) mixed with TTS-generated audio (Microsoft Neural TTS / edge-tts) across English, French, Spanish, and German.

Key design principle: Voicemail greetings are recorded by real humans and sound acoustically identical to live speech. This dataset includes real spoken content with semantic cues so models learn to distinguish them by WHAT is being said, not just HOW it sounds.

Classes

Label	ID	Description	Semantic Cues
human	0	Live human speech on phone	"Hello?", "Allô?", "¿Aló?", "Hallo?"
voicemail	1	Voicemail greeting + beep	"Leave a message", "Laissez un message", "Deja un mensaje"
ivr	2	IVR system (automated menus)	"Press 1 for...", "Tapez 1 pour...", "Pulse 1 para..."
answering_machine	3	Generic automated message + beep	"The person you are calling is not available..."

Statistics

• Total samples: 2,144 (train: 1,822 + test: 322)
• Audio format: 16kHz mono WAV, up to 10 seconds

By class

Class	Train	Test	Total
human	574	101	675
voicemail	417	73	490
ivr	415	74	489
answering_machine	416	74	490

By language

Language	Real Speech (MINDS14)	TTS Generated	Total
English (en)	200	~1,399	~1,599
French (fr)	75	~170	~245
Spanish (es)	50	~100	~150
German (de)	50	~100	~150

French has extra weight since it's the primary European market.

Data Sources

Real human speech

• PolyAI/MINDS14: Real callers to phone banking IVR systems
  • en-US, en-GB, en-AU: 200 samples
  • fr-FR: 75 samples
  • es-ES: 50 samples
  • de-DE: 50 samples

TTS-generated speech

• Microsoft Neural TTS (edge-tts): 47+ voices across 14 locales
• Voice style varies by class:
  • Human: fast, conversational, varied pitch
  • Voicemail: personal, warm, moderate pace + beep tone
  • IVR: slow, authoritative, monotone/robotic + optional DTMF
  • Answering Machine: generic, impersonal + long beep

Audio post-processing

• Telephony bandpass filter (300-3400 Hz)
• Background noise (SNR 18-35 dB)
• Random gain variation (0.6-1.3x)
• Voicemail: beep at 440-1000Hz appended
• IVR: 30% include DTMF response tones
• Answering Machine: beep at 1000-1400Hz + post-silence

Usage

from datasets import load_dataset, Audio

dataset = load_dataset("AbijahKaj/telephony-amd-dataset")
dataset = dataset.cast_column("audio", Audio(sampling_rate=16000))

sample = dataset["train"][0]
audio = sample["audio"]["array"]  # numpy float32 @ 16kHz
label = sample["label"]           # 0-3

Training

Designed for WhisperForAudioClassification (Whisper encoder + classification head):

pip install transformers datasets evaluate torch torchaudio soundfile accelerate
huggingface-cli login
python scripts/train_local.py

Scripts

• scripts/train_local.py — Full training script (5090-ready, CLI args)
• scripts/streaming_amd.py — Streaming inference module
• scripts/build_dataset.py — English dataset builder (real speech + TTS)
• scripts/add_multilingual.py — French/Spanish/German expansion script

Limitations

• TTS-generated speech may have subtle artifacts vs real recordings
• Limited template variety per language (~15-35 templates per class)
• Real speech only in "human" class (voicemail/IVR/AM are TTS-only, which is realistic since those ARE automated)
• No G.711 µ-law codec simulation (uses bandpass approximation)