freococo/myanmartts_generated_millions_audios

MyanmarTTS Generated Millions Audios

A large-scale Burmese (Myanmar) speech dataset containing over 7.02 million audio clips spanning 7,633+ hours, synthesized using MyanmarTTS — a from-scratch 31M-parameter flow-matching TTS model (pip install myanmartts).

The corpus is segmented into length-based subsets to facilitate curriculum learning, targeted length-bucketed training, and flexible streaming.

---

Dataset Summary

• Total Audio: 7,633.12 hours (~318 days)
• Total Clips / Rows: 7,029,426
• Total Parquet Files: 2,592 files
• Total Size: 2,258.24 GB (~2.26 TB)
• Audio Format: 44,100 Hz, 16-bit PCM WAV
• Duration Range: 0.3947 s – 10.0891 s

---

Dataset Structure & Subsets

The dataset is partitioned into subsets categorized primarily by sentence/character length buckets:

Subset Config	Parquet Files	Size (GB)	Clips / Rows	Duration (Hours)	Min Duration	Max Duration
audio_parquets_under20	32	18.79	158,069	63.49 hrs	0.3947 s	2.4381 s
audio_parquets_20_29	209	108.24	640,838	365.77 hrs	0.9868 s	3.4830 s
audio_parquets_30_39	369	252.08	1,116,995	851.99 hrs	0.9752 s	4.4930 s
audio_parquets_40_49	122	103.42	366,000	349.55 hrs	2.0318 s	5.3986 s
audio_parquets_50_59	707	726.33	2,120,344	2,455.10 hrs	1.9389 s	6.4435 s
audio_parquets_60_70	740	904.43	2,216,192	3,057.24 hrs	1.8924 s	7.6858 s
audio_parquets (general)	413	144.95	410,988	489.98 hrs	0.4760 s	10.0891 s
TOTAL	2,592	2,258.24	7,029,426	7,633.12 hrs	0.3947 s	10.0891 s

---

Data Fields

Each Parquet file contains approximately 5,000 rows with the following schema:

Column	Type	Description
id	int32	Unique global sentence identifier
text	string	Burmese transcript / text sentence
audio	Audio(sampling_rate=44100)	Audio dictionary containing decoded array or raw WAV bytes
sample_rate	int32	Audio sampling rate (44100)

---

Source Corpus

Synthesized text was curated from freococo/myanmar_spoken_corpus, which aggregates, cleans, and deduplicates Burmese web and spoken-style text from:

• FineWeb2
• DCAD2000
• Open web crawl corpora

---

How to Use

Given the large footprint (~2.26 TB), streaming mode (streaming=True) is strongly recommended unless you are targeting specific subsets or training on high-capacity storage clusters.

1. Stream the Full Dataset (Default)

from datasets import load_dataset

ds = load_dataset(
    "freococo/myanmartts_generated_millions_audios",
    split="train",
    streaming=True,
)

sample = next(iter(ds))
print("Text:", sample["text"])
print("Sample Rate:", sample["audio"]["sampling_rate"])  # 44100
print("Audio Array Shape:", sample["audio"]["array"].shape)

2. Load a Specific Length Subset

from datasets import load_dataset

# Load short sentences (< 20 chars)
short_clips = load_dataset(
    "freococo/myanmartts_generated_millions_audios",
    name="audio_parquets_under20",
    split="train",
    streaming=True,
)

sample = next(iter(short_clips))
print(sample["text"])

3. Load Without Automatic Audio Decoding (Faster Throughput)

If you need only the raw WAV bytes or text:

from datasets import load_dataset

ds = load_dataset(
    "freococo/myanmartts_generated_millions_audios",
    name="audio_parquets_30_39",
    split="train",
    streaming=True,
)
ds = ds.cast_column("audio", None)  # Avoids decoding overhead

sample = next(iter(ds))
raw_bytes = sample["audio"]["bytes"]

---

License

This dataset is released under the Creative Commons Zero 1.0 Universal (CC0 1.0) license — Public Domain Dedication. You are free to copy, modify, distribute, and perform the work, even for commercial purposes, without asking permission.