fleurs-mn / README.md
btsee's picture
Add dataset card
5c9b669 verified
metadata
language:
  - mn
license: cc-by-4.0
task_categories:
  - automatic-speech-recognition
task_ids: []
pretty_name: FLEURS  Mongolian (Clean)
tags:
  - mongolian
  - speech
  - audio
  - fleurs

FLEURS — Mongolian (Quality-Filtered)

A quality-filtered version of the FLEURS (mn_mn) Mongolian benchmark dataset, cleaned for use with oron-tts (F5-TTS / Flow Matching).

Source

Derived from google/fleurs config mn_mn. FLEURS is the speech version of the FLoRes machine translation benchmark, covering 2,009 n-way parallel sentences across 102 languages.

Cleaning Pipeline

6-stage automated quality filter, thresholds calibrated for Mongolian TTS training (low-resource language; DeepFilterNet denoising applied downstream in oron-tts):

Stage Method Threshold
1. Format normalization librosa mono · 16 kHz
2. Voice activity detection Silero VAD ≥25 % speech frames
3. SNR filter RMS-based SNR ≥8 dB
4. Pitch metadata CREPE F0 recorded when available; not a rejection gate
5. AI quality score DNSMOS P.835 OVR ≥2.2 · SIG ≥2.4 · BAK ≥2.0
6. Full sentence verification Whisper large-v3 + CER CER ≤0.35, or ≤0.50 when length ratio is 0.75–1.25

Ground truth for sentence verification: raw_transcription field. Clips are kept between 1–30 seconds to match oron-tts training limits. All passing clips are peak-normalized to −1 dBFS and resampled to 24 kHz.

Schema

All original FLEURS fields preserved, plus computed quality metrics:

Field Type Description
id int32 Sample ID
num_samples int32 Number of audio samples
path string Audio file path
audio Audio(24000) Cleaned audio resampled to 24 kHz
raw_transcription string Original (unnormalized) transcription
transcription string Normalized transcription
gender int32 Speaker gender class
lang_id int32 Language class ID
language string Language name
lang_group_id int32 Language group class ID
snr_db float32 SNR in dB
mean_f0_hz float32 Mean F0 (Hz)
pitch_confidence float32 CREPE pitch confidence
dnsmos_sig float32 DNSMOS signal quality
dnsmos_bak float32 DNSMOS background noise
dnsmos_ovr float32 DNSMOS overall MOS
dnsmos_p808 float32 DNSMOS P.808 MOS
cer float32 CER vs. raw_transcription
asr_transcript string Whisper large-v3 output
duration_s float32 Duration in seconds

Usage

from datasets import load_dataset
ds = load_dataset("btsee/fleurs-mn")
sample = ds["train"][0]

License

CC BY 4.0