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.

• 1,831,846 examples in the default config: one row per sample with voice, two ambient-background variants (v3, v4) and Fish Audio S2 Pro codes
• 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

All seven subsets below are registered as loadable configurations with a train split.

Config	Content	Parquet files
default	Unified single-speaker table: voice, references, ambient backgrounds v3/v4 and codec codes	371
multispeaker	Multispeaker synthetic speech with reference audio and text	537
music_sfx_step2	Multispeaker scenes with separate voice, music, SFX, background and mixed audio	1,053
dynamic_sfx_step2	Multispeaker scenes with dynamic SFX and separate audio stems	1,052
music_sfx_step3	Available Step 2 music scenes with regenerated backgrounds	80
dynamic_sfx_step3	Available Step 2 dynamic-SFX scenes with regenerated backgrounds	54
wavcaps_bg	Recorded BBC + Freesound backgrounds in one tagged table, including crowd/babble	216

File counts are not row counts. Some generation runs were stopped before exhausting their source data, so Step 3 should not be assumed to cover all rows in Step 2, and the v4 background is missing for part of default (see below). WavCaps publication is complete: 33,409 recordings, approximately 1,959 hours.

Load any subset

from datasets import load_dataset

ds = load_dataset(
    "mlinmg/translation_dataset", "multispeaker", split="train", streaming=True
)
# Replace "multispeaker" with any config name in the table above.

Multispeaker and background fields

multispeaker contains id (string), language, transcription, references (a list of reference audio bytes and text), and generated_audio (bytes/path). The default IDs are integers; do not assume all configs share an ID type.

The four music_sfx_step2/3 and dynamic_sfx_step2/3 configs additionally contain duration_seconds, n_speakers, bg_prompt, bg_profile, and separate generated_audio, sfx_audio, music_audio, bg_audio, mixed_audio columns. Their references also retain speaker_id, source, revision, and sha256.

wavcaps_bg has a distinct schema with source, tags, caption, and provenance; see its dedicated section below.

Structure (default)

One row per unique sample_unique_id. The table joins the former one_speaker, ambient_v3 and ambient_v4 subsets with the codec codes of mlinmg/codecs-set.

field	type	description
id	int64	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}
sample_unique_id	string	xxh64(f"{id}_{language}_{transcription}"), unique per row
voice_audio	Audio	synthesized clean speech
bg_prompt_v3, bg_profile_v3	string	prompt and profile of the v3 ambient background
bg_audio_v3	Audio	v3 ambient background
bg_prompt_v4, bg_profile_v4	string	prompt and profile of the regenerated v4 background, or null
bg_audio_v4	Audio	regenerated v4 ambient background, or null
voice_codes	list<list<int64>>	Fish Audio S2 Pro codes of voice_audio
reference_codes	list<struct>	Fish Audio S2 Pro codes of the references, {text, codes}

v4 generation was stopped before covering the whole corpus: 1,221,781 rows have a v4 background and 610,065 rows have all *_v4 fields null. Every row has a v3 background. Where a source subset contained several rows with the same sample_unique_id, only the first one was kept.

Multi-speaker configs add 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", split="train", streaming=True)  # default config
ex = next(iter(ds))

The default config alone is about 2 TB; 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. Publication is complete; wavcaps_bg/PUBLICATION_COMPLETE.json records the verified shards.

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")