TTS Loader

Loader for TTS (Text-to-Speech) datasets.

There are two parsing strategies for TTS datasets, controlled by the root_strategy field in the schema.

Strategies

Index-based (default)

A CSV / TSV / pipe-delimited index file maps audio paths to transcriptions.

Controlled by:

Field	Required	Description
format	✗	Optional file format hint ("csv", "tsv", "pipe"). When omitted, the loader infers it from index_file where possible.
index_file	✓	Path to the index file, relative to the dataset root.
base_audio_path	✗	Directory prefix or list of directories used to resolve file_path dtype columns.
columns	✗	Column mappings from source columns to logical names.
separator	✗	Explicit column separator (e.g. "\|").
has_header	✗	Whether the index file has a header row. When false, source_column must be a positional integer.
encoding	✗	File encoding (e.g. "utf-8-sig" for files with a BOM).

Paired-file / glob-based (root_strategy: "paired_glob")

Each audio file has a matching .txt file with the same stem. The loader recursively finds all text files, reads the transcription, and pairs them with the corresponding audio file. The parent directory name is captured as a split column in the resulting DataFrame.

Controlled by:

Field	Required	Description
file_pattern	✓	Glob pattern used to find text files (e.g. "**/*.txt").
audio_extension	✓	Extension of the matching audio files (e.g. ".webm").
content_mapping	✗	Optional mapping of file content to DataFrame columns.

---

Examples

Index-based schema

# Example: pipe-delimited, headerless metadata
dataset_id: "aso-ckb-tts"
task: "TTS"
format: "pipe"
separator: "|"
has_header: false
index_file: "metadata.csv"
base_audio_path: "wavs/"
columns:
  audio_path:
    source_column: 0        # positional index (no header)
    dtype: "file_path"
  transcription:
    source_column: 1
    dtype: "string"

If the index stores IDs instead of directly joinable filenames, the shared file_path resolver also supports search-based matching:

dataset_id: "example-tts"
task: "TTS"
index_file: "metadata.csv"
base_audio_path:
  - "wavs/"
  - "backup_wavs/"

columns:
  audio_path:
    source_column: "clip_id"
    dtype: "file_path"
    path_match_strategy: "exact"   # or "contains"
    file_extension: ".wav"
  transcription:
    source_column: "text"
    dtype: "string"

You can also build filenames and directory roots from metadata columns without hardcoding dataset-specific logic:

dataset_id: "example-tts-dynamic-path"
task: "TTS"
index_file: "metadata.tsv"
base_audio_path: "${split}/"

columns:
  audio_path:
    source_column: "sentence_id"
    dtype: "file_path"
    file_extension: ".wav"
    path_template: "${speaker_id}_khm_${value}"
  transcription:
    source_column: "text"
    dtype: "string"

For each row, that resolves to dataset_root / <split>/<speaker_id>_khm_<sentence_id>.wav.

Paired-glob schema

dataset_id: "pl-PL-darkman"
task: "TTS"
root_strategy: "paired_glob"
file_pattern: "**/*.txt"
audio_extension: ".webm"