TTS-Clean44k

A multilingual pool of verified-clean, wideband speech for training and evaluating speech restoration / text-to-speech (TTS) models. Every utterance is independently checked on two axes and stored as parquet with its per-utterance quality scores attached:

1. Native sample rate ≥ 44.1 kHz — measured per file with ffprobe, never trusting the source's advertised rate. Anything below 44.1 kHz is dropped.
2. DNSMOS P.835 bak ≥ 3.644 — the background-noise MOS from the DNSMOS P.835 model. Only genuinely clean recordings pass.

The dataset was assembled as the clean teacher pool for Sidon call-centre speech restoration (the decoder is trained to reproduce its teacher, so the teacher must be genuinely clean and full-band), but it is broadly useful as a filtered multilingual TTS corpus.

Composition

28 language/source configs · 119,950 utterances · 208.7 h · utterance-weighted DNSMOS bak 4.036 / sig 3.484 / ovrl 3.203.

#	config	utts	hours	bak	sig	ovrl
1	hifitts	19,447	30.0	4.026	3.533	3.247
2	uk_ireland_en	14,925	28.0	4.077	3.572	3.313
3	cv_zhcn	8,948	15.4	3.952	3.466	3.145
4	basque	6,907	13.5	4.043	3.521	3.230
5	galician	5,343	10.0	4.055	3.521	3.248
6	catalan	4,018	9.0	3.977	3.433	3.127
7	peruvian_es	4,991	8.8	4.052	3.423	3.154
8	south_african	5,240	8.3	3.977	3.460	3.149
9	kannada	3,542	7.6	4.045	3.436	3.165
10	gujarati	3,789	7.4	4.079	3.489	3.226
11	argentinian_es	4,362	6.8	4.077	3.520	3.246
12	colombian_es	4,105	6.8	4.052	3.404	3.144
13	chilean_es	3,755	6.6	4.056	3.424	3.160
14	tamil	3,633	6.4	4.027	3.300	3.035
15	nigerian_en	2,687	5.1	4.102	3.548	3.299
16	javanese	2,906	4.2	4.004	3.467	3.163
17	malayalam	2,470	4.1	4.034	3.367	3.095
18	aishell3	2,814	4.1	4.047	3.513	3.218
19	telugu	2,517	4.0	4.053	3.358	3.098
20	venezuelan_es	2,466	4.0	4.057	3.473	3.208
21	burmese	2,071	3.6	4.020	3.495	3.203
22	sundanese	2,116	3.5	3.999	3.443	3.140
23	khmer	2,098	3.2	4.092	3.448	3.202
24	marathi	1,400	2.8	4.088	3.433	3.186
25	yoruba	1,464	2.2	4.035	3.476	3.178
26	nepali	1,157	2.0	4.000	3.496	3.183
27	puertorico_es	516	0.9	4.050	3.443	3.169
28	cv_ta	263	0.4	3.855	3.116	2.780

Schema

Each config is a single train split with columns:

column	type	description
audio	Audio(sampling_rate=48000)	mono waveform, decoded to 48 kHz
source	string	source/config name
bak	float32	DNSMOS P.835 background-noise MOS (≥ 3.644 for every row)
sig	float32	DNSMOS P.835 signal MOS
ovrl	float32	DNSMOS P.835 overall MOS
duration	float32	clip length in seconds

Long recordings are chunked to ≤ 15 s; short utterances (≥ 4 s) are kept whole.

Usage

from datasets import load_dataset

# load one config (source)
ds = load_dataset("Scicom-intl/TTS-Clean44k", "uk_ireland_en", split="train")
print(ds[0]["audio"], ds[0]["bak"], ds[0]["duration"])

# or stream a large config
ds = load_dataset("Scicom-intl/TTS-Clean44k", "hifitts", split="train", streaming=True)

Sources

• OpenSLR high-quality TTS — Javanese (41), Sundanese (44), Tamil (65), Telugu (66), Malayalam (63), Marathi (64), Khmer (42), Nepali (43), Gujarati (78), Kannada (79), Burmese (80).
• OpenSLR crowdsourced — Argentinian/Chilean/Colombian/Peruvian/Puerto-Rican/Venezuelan Spanish (61/71/72/73/74/75), Catalan (69), Basque (76), Galician (77), Yoruba (86), Nigerian English (70), South-African English (32), UK & Ireland English (83).
• AISHELL-3 (Mandarin), Hi-Fi TTS (English, capped at 30 h for balance), Common Voice (Mandarin zh-CN, Tamil ta).

What the gates rejected

The verification is strict on purpose — sources that only claim to be high-fidelity were dropped:

• Common Voice id (Indonesian) and yue (Cantonese) — excluded entirely; every clip was below 44.1 kHz (Cantonese was uniformly 32 kHz).
• VCTK — omitted here only because its available mirror was download-throttled, not for quality.

Nothing below the sample-rate or DNSMOS bar is included.

Licensing

Audio is redistributed from the upstream corpora listed above; each retains its original license (OpenSLR corpora are variously CC BY / CC BY-SA / CC0, AISHELL-3 is research-use, Hi-Fi TTS is CC BY 4.0, Common Voice is CC0). Consult the corresponding source before commercial use. The DNSMOS scores and the 48 kHz re-encoding are provided as-is.