freococo/myanmartts_v3_full_arrows

MyanmarTTS v3 — Full Arrows (Training-Ready)

Preprocessed Arrow dataset derived from freococo/myanmartts_generated_millions_audios. All 7,029,426 source clips were converted to 24 kHz mono int16 PCM and packed into Arrow format.

Zero loss. Every source parquet row is accounted for. Source row count and arrow clip count both equal 7,029,426.

---

Dataset Summary

• Total Clips: 7,029,426
• Total Audio: 7,633.12 hours (~318 days)
• Total Arrow Batches: 31
• Total Arrow Size: 1,320.46 GB (~1.32 TB)
• Audio Format (in arrow): 24,000 Hz, mono, 16-bit PCM WAV bytes
• Duration Range: 0.3947 s – 10.0891 s
• Avg Clip Duration: 3.91 s

---

Subset Structure

Each subset is stored as one or more batches (batch_00000, batch_00100, …). Each batch folder contains:

• raw.arrow — the Arrow file
• duration.json — list of durations in seconds
• vocab.txt — tokenizer vocabulary (2,626 tokens)

Subset	Batches	Clips	Hours	Min (s)	Max (s)
audio_parquets_under20	1	158,069	63.49	0.3947	2.4381
audio_parquets_40_49	2	366,000	349.55	2.0318	5.3986
audio_parquets (general)	5	410,988	489.98	0.4760	10.0891
audio_parquets_20_29	3	640,838	365.77	0.9868	3.4830
audio_parquets_30_39	4	1,116,995	851.99	0.9752	4.4930
audio_parquets_50_59	8	2,120,344	2,455.10	1.9389	6.4435
audio_parquets_60_70	8	2,216,192	3,057.24	1.8924	7.6858
TOTAL	31	7,029,426	7,633.12	0.3947	10.0891

---

Arrow Schema

Each raw.arrow is written with pyarrow.ipc.new_file and has three columns:

Column	Type	Description
audio	struct<bytes: binary, path: string>	bytes holds the raw WAV (24 kHz, mono, int16). path is None.
text	string	Normalized Burmese transcript
duration	float32	Clip duration in seconds

---

Usage

Load with PyArrow

import pyarrow.ipc as ipc

with ipc.open_file("audio_parquets_under20/raw.arrow") as reader:
    table = reader.read_all()
    print(table.num_rows)  # 158,069
    print(table.schema)

Read Audio Bytes Directly

import io
import pyarrow.ipc as ipc
import soundfile as sf

with ipc.open_file("audio_parquets_under20/raw.arrow") as reader:
    batch = reader.get_batch(0)
    row = batch.slice(0, 1).to_pylist()[0]

wav_bytes = row["audio"]["bytes"]
data, sr = sf.read(io.BytesIO(wav_bytes))
print(sr)                    # 24000
print(data.shape)            # (samples,)
print(row["text"])           # Burmese transcript
print(row["duration"])       # seconds

Merge All Batches

Each batch is self-contained. To combine them, read the record batches from every raw.arrow and write them into a single output arrow. All 31 batches share the same schema.

---

Verification

• Source Parquet Audit: Independent footer scan of all 2,592 source parquet files yielded 7,029,426 rows — matches HF-reported row count.
• Arrow Output Count: Arrow clip count aggregate from every duration.json file equals 7,029,426 clips.
• Delta: 0 (no clips skipped, no clips added).
• Sampling Tests: Confirmed sample_rate=24000, duration matches actual audio length, and text decodes cleanly as valid Burmese script.

---

Source

• Original 44.1 kHz Parquet Dataset: freococo/myanmartts_generated_millions_audios (2,258.24 GB, 2,592 parquet files)
• Base TTS Model: MyanmarTTS

---

License

Released under CC0 1.0 Universal (CC0 1.0) — Public Domain Dedication.