ZeroSpeech

A large synthetic Vietnamese speech corpus for ASR training: 9,867,987 utterances / 26,896 hours, spoken by 199,265 distinct voices, generated with ZeroTTS from web and conversational text.

Every clip is 16 kHz mono FLAC, 1–30 s, paired with the exact text it was synthesized from.

Fields

field	type	description
audio	Audio(16 kHz)	the waveform, FLAC-encoded
text	string	the transcript — the exact string given to the TTS
source	string	which corpus the text came from: vietvault (web documents) or blossoms (conversational turns)
kind	string	vi (Vietnamese), mixed (Vietnamese with embedded English), en (English only)
duration	float32	seconds, measured from the audio
id	string	stable clip id
voice_id	int32	which reference voice spoke it
speaker_id	int32	source-corpus speaker label, -1 where unknown
voice_lang	string	the language the reference voice was recorded speaking

from datasets import load_dataset

ds = load_dataset("zeroweight-ai/ZeroSpeech", split="train", streaming=True)
row = next(iter(ds))
row["audio"]["array"], row["text"], row["source"]

Composition

	clips
from vietvault (web)	4,963,772
from blossoms (conversational)	4,904,215
total	9,867,987

text type	clips	share
vi — Vietnamese	5,104,533	51.7 %
mixed — Vietnamese + English	4,604,354	46.7 %
en — English only	159,100	1.6 %

Code-switched text is kept deliberately and is a large share of the corpus: it is what a naive language-ID filter throws away and what Vietnamese ASR most often gets wrong.

duration	clips
1–3 s	650,978
3–5 s	1,439,713
5–8 s	2,495,878
8–12 s	2,393,627
12–18 s	1,799,037
18–24 s	824,950
24–31 s	254,294

Mean duration 9.8 s. Voices: 199,265 distinct references, 29,448 of them carrying an exact speaker label.

How it was made

Text was drawn from vietvault web documents and BlossomsAI/vietnamese-conversational-dataset chat turns, then markdown-stripped, sentence-segmented, filtered for speakable content (character whitelist, real casing, terminal punctuation, digit-density and repetition limits) and packed into utterances targeting a 1–30 s length distribution.

Each utterance was then synthesized by ZeroTTS — a zero-shot Vietnamese TTS model — conditioned on a reference voice drawn at random from a bank of 199,265 speaker references and matched to the utterance's language. No text normalization frontend was applied: the model received raw orthography.

Generated clips were dropped if the model never emitted its stop token, if the measured duration diverged from the character-rate estimate (the signature of a repetition loop), or if the output was near-silent.

Limitations — read before training on this

• The audio is synthetic. Every clip comes from a single model (ZeroTTS), so it carries that model's phonetics, prosody and artifacts throughout. Acoustic conditions do vary — the reference voices are real recordings spanning a wide range of rooms, channels and noise floors, and the model reproduces those conditions — but the speech itself is read-aloud in character: no spontaneous disfluency, no overlapping speakers, no genuine conversational timing. Useful as pretraining or augmentation mass; fine-tune on real speech.
• Text normalization is the TTS model's. Digits and acronyms were fed as orthography, so however the model read 2/7 or TT-BGTVT is what the transcript claims was said.
• Source text is web-scraped and carries whatever bias, error and distributional skew the source corpora do.

Related

• zeroweight-ai/ZeroTTS — the zero-shot TTS model that generated every clip here.

License

Released under CC-BY-NC-4.0. The source text corpora carry their own licenses; check them before commercial use.