mcamara/vtl-speech-landmarks

VTL Speech Landmarks Dataset

Articulatory speech synthesis dataset with acoustic landmarks, generated using VocalTractLab (VTL).

Dataset Description

This dataset contains synthesized speech for 117,497 English words from the CMU Pronouncing Dictionary, generated with two speakers (male and female). Each word includes:

• Audio: 48kHz WAV files
• Landmarks: Acoustic-phonetic event markers (JSON)
• Articulatory data: Full vocal tract trajectories from VTL (JSON)

Speakers

Speaker	Base F0	Emphasis
Male	120 Hz	1.2
Female	200 Hz	1.4

Dataset Structure

vtl-speech-landmarks/
├── male/
│   ├── wav/                     # 117,497 audio files
│   ├── landmarks/               # 117,497 landmark JSON files
│   └── articulatory/            # 117,497 articulatory JSON files
└── female/
    ├── wav/                     # 117,497 audio files
    ├── landmarks/               # 117,497 landmark JSON files
    └── articulatory/            # 117,497 articulatory JSON files

Landmark Types

Type	Name	Description
V	Vowel	Maximum mid-frequency energy
G	Glide	Energy change in formant region (for r, l, w, y)
Sc	Stop Closure	Start of oral closure
Sr	Stop Release	Burst energy when stop releases
Fc	Fricative Closure	Peak frication turbulence onset
Fr	Fricative Release	Transition out of fricative
Nc	Nasal Closure	Abrupt change when nasal begins
Nr	Nasal Release	Abrupt change when nasal ends

File Formats

Landmarks JSON (*_landmarks.json)

{
  "word": "hello",
  "pronunciation": "HH AH0 L OW1",
  "arpabet": ["HH", "AH0", "L", "OW1"],
  "vtl_phonemes": ["h", "@", "l", "O", "U"],
  "speaker": "male",
  "duration_ms": 520.0,
  "sample_rate": 48000,
  "landmarks": [
    {"type": "Fc", "time_ms": 50.0, "phoneme": "h", "ipa": "h", "confidence": 0.95},
    {"type": "V", "time_ms": 180.0, "phoneme": "@", "ipa": "ʌ", "confidence": 0.92},
    ...
  ],
  "phoneme_timings": [
    {"phoneme": "h", "start": 0.05, "end": 0.12, "duration": 0.07},
    ...
  ]
}

Articulatory JSON (*_articulatory.json)

{
  "sample_rate": 400.0,
  "num_frames": 208,
  "time": [0.0, 0.0025, 0.005, ...],
  "velum_opening": [...],
  "lip_aperture": [...],
  "tongue_tip_constriction": [...],
  "tongue_body_constriction": [...],
  "tube_areas": [[...], [...], ...]
}

Audio (WAV)

• Format: 16-bit PCM
• Sample rate: 48,000 Hz
• Channels: Mono

Usage

Download specific files

from huggingface_hub import hf_hub_download

# Download a landmark file
landmarks_path = hf_hub_download(
    repo_id="mcamara/vtl-speech-landmarks",
    filename="male/landmarks/hello_landmarks.json",
    repo_type="dataset"
)

# Download audio
audio_path = hf_hub_download(
    repo_id="mcamara/vtl-speech-landmarks",
    filename="male/wav/hello.wav",
    repo_type="dataset"
)

Load landmarks

import json

with open(landmarks_path) as f:
    data = json.load(f)

print(f"Word: {data['word']}")
print(f"Duration: {data['duration_ms']} ms")
for lm in data['landmarks']:
    print(f"  {lm['type']} at {lm['time_ms']:.1f} ms - {lm['phoneme']} ({lm['ipa']})")

Load and play audio

import soundfile as sf

audio, sr = sf.read(audio_path)
print(f"Sample rate: {sr}, Duration: {len(audio)/sr:.2f}s")

Generation Details

• Source: CMU Pronouncing Dictionary (117,497 words)
• Synthesizer: VocalTractLab (VTL) articulatory speech synthesizer
• Landmark extraction: Energy-based detection from spectral analysis
• Phoneme mapping: ARPABET to VTL phoneme conversion with diphthong expansion

Applications

• Acoustic-phonetic research
• Speech recognition training data
• Text-to-speech development
• Phonetic landmark detection models
• Articulatory synthesis research

Citation

If you use this dataset, please cite:

@misc{vtl-speech-landmarks,
  title={VTL Speech Landmarks Dataset},
  year={2025},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/mcamara/vtl-speech-landmarks}}
}

License

MIT License