translation_dataset

Synthetic, expressive, multilingual speech for cross-lingual dubbing research. Each example pairs a style-annotated text with generated audio that clones an English reference voice: the voice stays the same, the language changes.

• ~1.9M examples in the one_speaker config
• 17 languages
• ~900 distinct reference speakers

Samples

Each sample shows the generated audio followed by the English reference voice that conditioned it.

English (en)

Italian (it)

Russian (ru)

Arabic (ar)

Chinese (zh)

Japanese (ja)

Korean (ko)

Hindi (hi)

Configs

config	content
one_speaker	single-speaker voice only
ambient_v3, ambient_v4	single-speaker voice with ambient background
music_sfx_step2, dynamic_sfx_step2	multi-speaker scenes with separate voice, SFX and music stems
music_sfx_step3, dynamic_sfx_step3	same scenes as Step 2 with regenerated backgrounds

Some configs are still being uploaded and may be incomplete.

Structure (one_speaker)

field	type	description
id	int	utterance ID; the same id appears once per language, linking parallel translations
language	string	ISO 639-1 code
transcription	string	text with inline style directives
references	list<struct>	reference voices, {audio, text}
generated_audio	Audio	synthesized speech

Background configs add fields such as bg_prompt, bg_profile, bg_audio, and, for multi-speaker configs, separate sfx_audio, music_audio and mixed_audio stems.

Style directives

Transcriptions contain bracketed prosodic cues that are meant to be performed, not read:

[inhale] So, the detective... [soft voice] just looked at the file
[long pause] and then he [gasp] he said it. [angry] It w-w-wasn't a robbery.
[shouting] It was an ambush!!!

Stuttering, syllabic emphasis and repeated punctuation are also used as intensity cues. These cues are requests to the synthesizer and are not verified perceptual annotations.

Loading

from datasets import load_dataset

ds = load_dataset("mlinmg/translation_dataset", "one_speaker", split="train", streaming=True)
ex = next(iter(ds))

The dataset is several hundred GB; streaming is recommended.

Limitations

• Synthetic configs: their audio is synthetic and should not be treated as recorded human speech. wavcaps_bg contains recorded-source assets; see its section below.
• English-only reference voices, so target-language accents may be influenced by the source timbre.
• Intelligibility, speaker fidelity and background quality have not been exhaustively validated. Synthesis artifacts may be present.
• License to be determined.

Recorded WavCaps backgrounds (wavcaps_bg)

One combined train table containing BBC and Freesound background candidates from cvssp/WavCaps at revision 0930ec11ded28fa0eaa910fde2f6fc3538acbeac. These are recorded-source assets, distinct from the synthetic configs above. Uploads may be partial until wavcaps_bg/PUBLICATION_COMPLETE.json exists.

bg_audio contains the original encoded audio, with no resampling or normalization. id is prefixed by source; source is bbc or freesound. Full source metadata and original IDs are retained in source_metadata_json. duration is measured in seconds; sha256 hashes the original encoded audio; source_revision pins provenance.

Filter using the list-valued tags column: nature, urban, indoor_mechanical, crowd (including chatter/babble), ambience, and music_mentioned. Tags come from source captions/metadata, not an acoustic classifier. Crowd, babble and environmental music are allowed. Candidates are at least 15 seconds long. Decoded audio integrity and byte preservation are checked, but this does not certify perceptual quality or speech-free backgrounds: acoustic_background_validation is false for every row.

from datasets import load_dataset
ds = load_dataset("mlinmg/translation_dataset", "wavcaps_bg", split="train", streaming=True)
crowd = ds.filter(lambda row: "crowd" in row["tags"])
bbc = ds.filter(lambda row: row["source"] == "bbc")