freococo/myanmar_bible_voices

🇲🇲 Myanmar Bible Speech 44k Corpus

A high-quality, sentence-level segmented Myanmar (Burmese) speech dataset containing 44,371 utterances (~60.1 hours) extracted from 917 Bible chapters.

All audio clips are aligned at the word and sentence level using Meta's MMS (Massively Multilingual Speech) Forced Aligner (mms-300m-1130-forced-aligner) and tokenized using mmdt-tokenizer.

---

📊 Dataset Summary

• Total Utterances: 44,371
• Total Duration: ~60.06 hours
• Average Utterance Length: 4.87 seconds
• Audio Format: 16,000 Hz, Mono, 16-bit PCM WAV (Embedded in Parquet)
• Language: Myanmar / Burmese (my / mya)
• Number of Chapters: 917 complete books and chapters

---

📑 Field Descriptions

Column	Type	Description	Example
audio	Audio	Embedded 16kHz mono audio stream (with built-in web player)	[Audio Player]
file_name	string	Continuous 5-digit filename identifier	00001.wav
transcript_original	string	Exact original Burmese source text	ကမ္ဘာဦးကျမ်း အခန်းကြီး (၁)
transcript_char	string	Normalized text where numbers are verbalized to spoken Burmese words	ကမ္ဘာဦးကျမ်း အခန်းကြီး (တစ်)
transcript_tokenised	string	Syllable-tokenized units from original text via mmdt-tokenizer	ကမ္ဘာ ဦး ကျမ်း အ ခန်း ကြီး ( ၁ )
duration	float64	Audio clip length in seconds	4.36

---

🚀 Quickstart & Usage

1. Standard Loading

from datasets import load_dataset

# Load entire dataset
dataset = load_dataset("freococo/myanmar_bible_speech_44k", split="train")

# Access first sample
print(dataset[0])
# Play audio
# dataset[0]["audio"]["array"]

2. Streaming Mode (Instant, No Download Wait)

from datasets import load_dataset

# Stream samples on demand without downloading the full ~7.5 GB
stream_ds = load_dataset("freococo/myanmar_bible_speech_44k", split="train", streaming=True)

for sample in stream_ds:
    print(sample["file_name"], "->", sample["transcript_char"])
    break

---

🛠️ Data Processing Pipeline

1. Audio Preprocessing: 917 master chapter recordings were converted to 16kHz mono audio. Corrupt frames were repaired using FFmpeg frame sanitization.
2. Sentence Splitting: Transcripts were partitioned on the Burmese section sign (။), psalm headings, and comma clauses (၊) ensuring sentence lengths under 200 characters.
3. Forced Alignment: Forced alignment executed using Meta's mms-300m-1130-forced-aligner with uroman phonetic mapping on GPU.
4. Audio Slicing: Microsecond NumPy array slicing applied with a 60ms acoustic boundary safety pad to prevent syllable clipping.
5. Text Verbalization & Syllable Tokenization:
   • Numbers and digits mapped to spoken Burmese phonological words using recursive Myanmar numeral rules.
   • Syllables tokenized via mmdt-tokenizer.