Phitran21/Vietnamese-Adam-TTS-50h

Vietnamese Adam TTS 50h

A Vietnamese single-speaker synthetic speech dataset featuring the Adam male voice, intended for training and evaluating text-to-speech (TTS) systems.

Dataset Overview

Property	Value
Language	Vietnamese (vi)
Speaker	Adam (male voice preset)
Audio samples	18,478
Total duration	50.5463 hours (approximately 50 hours, 32 minutes, 47 seconds)
Average clip duration	Approximately 9.85 seconds
Sample rate	48,000 Hz
Audio format	WAV, mono, 16-bit PCM
Speech type	Synthetic, single-speaker
Archive	finnal_tts.zip
Archive size	Approximately 17.5 GB (reported archive size)
Primary task	Text-to-Speech

Duration and sample count are based on the generation cache: 18,478 completed entries and the sum of their stored clip durations. The total has not been independently verified by decoding every WAV file.

Dataset Structure

The dataset is distributed as a single archive, finnal_tts.zip. Once extracted:

data.jsonl
wav/
  1.wav
  2.wav
  3.wav
  ...

data.jsonl contains one JSON object per line with numeric id and Vietnamese text fields, for example:

{"id": 1, "text": "Ví dụ một câu tiếng Việt."}

Audio clips are stored under wav/<id>.wav, where <id> matches the corresponding numeric id in data.jsonl. The example text above illustrates the format and is not a quoted dataset entry.

Text Normalization

All Vietnamese text in data.jsonl has been pre-normalized before speech generation.

• Numbers and abbreviations have been converted into their spoken Vietnamese forms.
• Text is provided in a TTS-ready format.
• No additional text normalization is required for basic dataset usage.
• The original normalized text is preserved and paired with its corresponding WAV file.

Note: Users may still apply model-specific tokenization or phonemization when training.

Download

Download the complete dataset — finnal_tts.zip

The archive is large (approximately 17.5 GB). Ensure sufficient free disk space for both the ZIP archive and the extracted WAV files.

Intended Uses

This dataset can be used for training, fine-tuning, or evaluating Vietnamese speech synthesis systems, including architectures such as Matcha-TTS / flow matching, VITS / Piper, FastSpeech 2, and custom TTS models. Suitability depends on the preprocessing and training setup.

Data Quality and Limitations

• All speech is synthetically generated using the Adam voice preset.
• Clip-level technical checks were performed during dataset generation; unsuccessful generations were retried.
• Text transcription, pronunciation, prosody, and speaker consistency have not been manually verified for every clip.
• The reported duration is derived from generation records, not from an independent full-file audit.
• This dataset is not a recording of a real human speaker.

License

The dataset is labeled CC BY 4.0. Users should independently verify that their intended use complies with the terms of any upstream speech-generation model, voice assets, and data sources.