Datasets:

espnet
/

Bagpiper_TTS_SFT_Data

Every row contains one complete natural-language-guided synthesis example and its encoded audio. Parquet shards target approximately 256 MiB and never split an example.

Field	Type	Description
`row_id`	string	Globally unique `source_subset:example_id` release key
`example_id`	string	Original source identifier
`partition`	string	One of the six public application names
`source_subset`	string	Internal source corpus retained for provenance
`split`	string	Release split, initially `train`
`direction`	string	`text_to_audio`
`messages`	list of structs	Ordered `{role, modality, content, audio_index}` conversation
`audio`	struct	`{path, bytes}` containing the complete encoded target audio
`audio_sha256`	string	SHA-256 of encoded audio bytes
`audio_bytes`	int64	Encoded byte length
`audio_format`	string	Source container/codec suffix
`sample_rate`	int32	Decoded sample rate
`channels`	int16	Channel count
`num_frames`	int64	Decoded frame count
`duration_seconds`	float64	Decoded duration
`metadata_json`	string	Sanitized source annotations; no paths or usernames
`source_revision`	string	Pinned raw Hugging Face source revision
`schema_version`	string	`bagpiper-self-contained-parquet-v1`

Text is preserved verbatim. The audio-message content is replaced with a portable display filename and points to the embedded audio value by audio_index; no private source path is retained.

All rows use exactly this message pattern and contain no system role: user:text → assistant:text → assistant:audio. Every row contains exactly one audio object. Intended reuse of valid audio across rows is preserved.