mcamara/all-words-in-english-with-pink-trombone

Dataset Card for Pink Trombone English Phonetic & Landmark Dataset

Repository: mcamara/all-words-in-english-with-pink-trombone Modality: Audio + time-aligned events (landmarks) + articulatory keyframes Language: English (IPA) Sampling rate: 44,100 Hz (mono) Voices: two synthetic voices — M (male) and F (female)

---

Summary

A large-scale, clean synthetic speech dataset generated with the Pink Trombone articulatory synthesizer. Every English dictionary word is synthesized in two voices (male and female). Each example links a word to:

1. audio — the synthesized waveform (44.1 kHz, mono, FLAC — lossless),
2. utterance — the articulatory keyframes driving the synthesis ({name, keyframes}),
3. landmarks — time-aligned acoustic landmarks detected from the audio.

Fields

Field	Type	Description
id	string	Orthographic word (e.g. "hello"). Repeats once per voice.
audio	Audio (44.1 kHz)	Synthesized mono waveform (FLAC, lossless).
utterance	string (JSON)	getUtterance() output: { "name", "keyframes": [...] } with per-phoneme articulatory parameters (tongue/constriction positions, tenseness, intensity, frequency, timing).
landmarks	string (JSON)	Array of { "type", "time", "name" }. Landmark times are in seconds.
sex	string	"M" (male) or "F" (female).

Landmark types

Type	Meaning
Sc / Sr	Stop closure / release
Fc / Fr	Fricative closure / release
Nc / Nr	Nasal closure / release
V	Vowel (mid-frequency energy peak)
G	Glide (formant transition)

Voice parameters

Voice	F0	Vocal tract length
M (male)	140 Hz	44
F (female)	220 Hz	38

Generation

Produced with the Pink Trombone web synthesizer driven headlessly via Playwright (pink-trombone-demos/batch-generator). For each word: text → IPA → articulatory keyframes (TTS module) → real-time synthesis + recording (Pink Trombone module) → WAV + landmark extraction (LEXI module). Landmarks are detected from energy/articulatory events. Audio is the synthesizer's native 44.1 kHz output (no resampling), stored as lossless FLAC.

Loading

from datasets import load_dataset
ds = load_dataset("mcamara/all-words-in-english-with-pink-trombone", split="train")
ex = ds[0]
ex["audio"]      # {'array': ..., 'sampling_rate': 44100}
ex["id"], ex["sex"]
import json
json.loads(ex["landmarks"])
json.loads(ex["utterance"])

# filter one voice
male = ds.filter(lambda r: r["sex"] == "M")

License

MIT.