NTQAI/nxkmer-tts-450

NXKMER TTS 450

Khmer speech recordings, transcripts, and speaker identifiers for text-to-speech research.

Dataset

The published train, validation, and test assignments and transcript text are preserved. Source audio is 48 kHz, mono, 16-bit PCM. The audio is encoded as lossless FLAC and embedded in Parquet shards; every recording is decoded after conversion and checked for exact equality with the source PCM samples. There is no resampling or trimming.

Split	Recordings	Hours	Speakers
train	101,316	454.292	38
validation	200	1.050	10
test	200	0.599	10

Total: 101,716 recordings, 455.941 hours, 38 speakers. The repository name uses a rounded duration.

Loading

This repository is public and uses automatic access requests. Request access on the dataset page, then authenticate with your approved account (hf auth login).

from datasets import load_dataset

dataset = load_dataset("NTQAI/nxkmer-tts-450", token=True)
example = dataset["train"][0]
print(example["text"], example["speaker"])

Streaming avoids downloading the full corpus:

from datasets import Audio, load_dataset

train = load_dataset("NTQAI/nxkmer-tts-450", split="train", streaming=True, token=True)
train = train.cast_column("audio", Audio(decode=False))
example = next(iter(train))
audio_bytes = example["audio"]["bytes"]  # Original-rate lossless FLAC

Recent versions of datasets use TorchCodec for automatic audio decoding. The Audio(decode=False) example exposes the embedded FLAC bytes and does not require automatic decoding.

Fields

• audio: embedded lossless FLAC bytes and a recording filename.
• text: unchanged Khmer transcript from the published metadata.
• speaker: unchanged source speaker identifier.
• audio_file: unchanged original WAV path from the published CSV; this is a provenance field. Load the recording from audio.
• duration_seconds, num_frames: exact source audio duration and frame count at 48,000 Hz.
• source_wav_sha256: checksum of the original WAV bytes.
• pcm_sha256: checksum of decoded PCM samples as little-endian signed 16-bit bytes.

License

Creative Commons Attribution 4.0 International (CC BY 4.0).

Paper citation