v1.0.0: pretrained model

This is TIFA 1.0 pretrained model.

Datasets

Voice datasets

• ~68h of private data with manual phoneme timing labels
• ~142h of public data with transcripts
  • Children's Song Dataset (English subset)
  • GTSinger (Chinese, English, and Japanese subsets, including paired speech)
  • MIR-1K (vocal channel)
  • OpenSinger
  • Opencpop
  • PopCS
  • M4Singer
  • No.7 Singing Database

Natural noise & accompaniments datasets

• DEMAND
• MUSAN (speech part excluded)
• MIR-1K
• MusicNet
• MUSDB18-HQ
• private Asian ethnic instruments dataset

Reverb dataset

MB-RIRs

Supported languages and special tags

Language	Code	Accepted forms	Quality of coverage
Chinese Mandarin	zh	Hanzi, pinyin	Excellent
English	en	Words	Good
Japanese	ja	Kanji, kana, romaji	Good
Yue (Cantonese)	yue	Hanzi, jyutping	Moderate

Tag	Meaning
AP	Aspiration
EP	Expiration
GS	Glottal stop

Specifications

• #Params: ~40M
• Max context length: ~1 minute (longer inputs may result in unstable alignment)

Acknowledgements

Data labels are provided by:

• 挚彬Club
• YQ之神
• 小狼
• 幽寂
• Kakaru
• 白烁

License

The model files apply CC BY-NC-SA 4.0 license. This license only constraints the model file itself, not the data that flows through it.