Dataset Viewer
Auto-converted to Parquet Duplicate
id
string
lang_id
int32
phonemes
string
mel_shape
list
mel
unknown
Akuapem_Twi_twi_00041924
0
balˈak bˈo fˈuw bˌalaˈam.
[ 80, 86 ]
[ 147, 78, 85, 77, 80, 89, 1, 0, 118, 0, 123, 39, 100, 101, 115, 99, 114, 39, 58, 32, 39, 60, 102, 52, 39, 44, 32, 39, 102, 111, 114, 116, 114, 97, 110, 95, 111, 114, 100, 101, 114, 39, 58, 32, 70, 97, 108, 115, 101, 44,...
Akuapem_Twi_twi_00039298
0
ˈama nˈe fapˈem hˈo ˈada hˈo.
[ 80, 86 ]
[ 147, 78, 85, 77, 80, 89, 1, 0, 118, 0, 123, 39, 100, 101, 115, 99, 114, 39, 58, 32, 39, 60, 102, 52, 39, 44, 32, 39, 102, 111, 114, 116, 114, 97, 110, 95, 111, 114, 100, 101, 114, 39, 58, 32, 70, 97, 108, 115, 101, 44,...
Akuapem_Twi_twi_00021276
0
ˌasoɾedˈan no ˌanuonjˈam.
[ 80, 86 ]
[ 147, 78, 85, 77, 80, 89, 1, 0, 118, 0, 123, 39, 100, 101, 115, 99, 114, 39, 58, 32, 39, 60, 102, 52, 39, 44, 32, 39, 102, 111, 114, 116, 114, 97, 110, 95, 111, 114, 100, 101, 114, 39, 58, 32, 70, 97, 108, 115, 101, 44,...
Akuapem_Twi_twi_00024571
0
nˈanso woɾˈenhu wˈon.
[ 80, 86 ]
[ 147, 78, 85, 77, 80, 89, 1, 0, 118, 0, 123, 39, 100, 101, 115, 99, 114, 39, 58, 32, 39, 60, 102, 52, 39, 44, 32, 39, 102, 111, 114, 116, 114, 97, 110, 95, 111, 114, 100, 101, 114, 39, 58, 32, 70, 97, 108, 115, 101, 44,...
Akuapem_Twi_twi_00032129
0
nˈa wopˈam no fˈi wiˈase.
[ 80, 86 ]
[ 147, 78, 85, 77, 80, 89, 1, 0, 118, 0, 123, 39, 100, 101, 115, 99, 114, 39, 58, 32, 39, 60, 102, 52, 39, 44, 32, 39, 102, 111, 114, 116, 114, 97, 110, 95, 111, 114, 100, 101, 114, 39, 58, 32, 70, 97, 108, 115, 101, 44,...
Akuapem_Twi_twi_00025892
0
nokwˈaɾe mmˌuadˈadi.
[ 80, 86 ]
"k05VTVBZAQB2AHsnZGVzY3InOiAnPGY0JywgJ2ZvcnRyYW5fb3JkZXInOiBGYWxzZSwgJ3NoYXBlJzogKDgwLCA4NiksIH0gICA(...TRUNCATED)
Akuapem_Twi_twi_00001506
0
nˈa sˈimɾi wˈoo mˈosa.
[ 80, 86 ]
"k05VTVBZAQB2AHsnZGVzY3InOiAnPGY0JywgJ2ZvcnRyYW5fb3JkZXInOiBGYWxzZSwgJ3NoYXBlJzogKDgwLCA4NiksIH0gICA(...TRUNCATED)
Akuapem_Twi_twi_00000577
0
hˈeles wˈoo ˌeleˈasa.
[ 80, 86 ]
"k05VTVBZAQB2AHsnZGVzY3InOiAnPGY0JywgJ2ZvcnRyYW5fb3JkZXInOiBGYWxzZSwgJ3NoYXBlJzogKDgwLCA4NiksIH0gICA(...TRUNCATED)
Akuapem_Twi_twi_00047081
0
ˈem ɪdʒˈɛktɪv a tˈe a ˈem ˈef o anˈim.
[ 80, 86 ]
"k05VTVBZAQB2AHsnZGVzY3InOiAnPGY0JywgJ2ZvcnRyYW5fb3JkZXInOiBGYWxzZSwgJ3NoYXBlJzogKDgwLCA4NiksIH0gICA(...TRUNCATED)
Akuapem_Twi_twi_00011993
0
asˈase sˈo nnˈipa njinˈaa.
[ 80, 86 ]
"k05VTVBZAQB2AHsnZGVzY3InOiAnPGY0JywgJ2ZvcnRyYW5fb3JkZXInOiBGYWxzZSwgJ3NoYXBlJzogKDgwLCA4NiksIH0gICA(...TRUNCATED)
End of preview. Expand in Data Studio

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

Ghana Speech Prepped (Filtered)

Precomputed mel spectrograms + phonemized text for multilingual Matcha-TTS training, filtered to clips in [1.0, 10.0] seconds and capped at 10 hours per language.

Stats

Metric Value
Train clips 485,575
Val clips 1,344
Total train audio 389.4 hours
Languages 42
Mel mean -5.277032676706492
Mel std 3.064793980356
Duration filter [1.0, 10.0] seconds
Cap 10 hours per language

Format

Each Parquet shard contains:

Column Type Description
id string Clip identifier (wav stem)
lang_id int32 Language ID (0-41), repurposes speaker slot
phonemes string LFN phonemized text
mel_shape list[int32] Shape of the mel spectrogram
mel binary Raw numpy bytes of the (un-normalized) mel spectrogram

Reconstruction

import io, json, numpy as np
from pathlib import Path
from datasets import load_dataset
from huggingface_hub import hf_hub_download

# Download metadata
stats = json.loads(hf_hub_download("ghananlpcommunity/ghana-speech-prepped-filtered", "ghana_speech_filtered.json", repo_type="dataset"))
lang_map = json.loads(hf_hub_download("ghananlpcommunity/ghana-speech-prepped-filtered", "lang_map.json", repo_type="dataset"))

# Load parquet shards
ds = load_dataset("ghananlpcommunity/ghana-speech-prepped-filtered")

# Reconstruct mel directory + filelists
out_dir = Path("ghana_speech_data")
mels_dir = out_dir / "mels"
mels_dir.mkdir(parents=True, exist_ok=True)

train_lines, val_lines = [], []
for split, lines in [("train", train_lines), ("val", val_lines)]:
    for row in ds[split]:
        mel = np.load(io.BytesIO(row["mel"]))
        np.save(mels_dir / f"{row['id']}.npy", mel)
        wav_path = str(out_dir / "wavs" / f"{row['id']}.wav")
        lines.append(f"{wav_path}|{row['lang_id']}|{row['phonemes']}")

(out_dir / "train_filtered.txt").write_text("\n".join(train_lines))
(out_dir / "val_filtered.txt").write_text("\n".join(val_lines))

Source

  • Original dataset: ghananlpcommunity/ghana-speech (42 language subsets, ~1.41M clips)
  • Phonemization: espeak lfn voice via twi_cleaners
  • Mel: n_fft=1024, n_mels=80, sample_rate=22050, hop_length=256, win_length=1024, f_min=0, f_max=8000
Downloads last month
289