Survey of Model C (TTS)

[wkCircle]

Survey of Model C, or Text to Speech model. Also survey which library is appropriate to manage model C. When the survey is done, please feel free to update README.md or upload files to conclude your results.

[wkCircle]

When the survey is done, please feel free to update README.md or upload files to conclude your results.

[wkCircle]

TTS resources:

General:

• TTS Intro
• What is TTS (Suno Bark, MS SpeechT5, MMS TTS)

Open AI TTS API:

• Pricing: Model	Usage
  Whisper	$0.006 / minute (rounded to the nearest second)
  TTS	$0.015 / 1K characters
  TTS HD	$0.030 / 1K characters

• API Tutorial

Speech-T5-TTS:

• Interactive Demo: only support for English, and the generated sound is too monotonic without cadence.
• Currently the English TTS are good for most models that I have tried. The problem is that when it comes to Chinese-English mixed text, it is hard for model to synthesize the output voice

Suno Bark Github, also supported Coqui TTS, or Hugging Face Framework

• Tutorial
• Bark Docs
• Bark supported preset voices: Bark Speaker Library v2. It supports multi-language, including simplified Chinese, but not Taiwanese tones. Users can click this speaker library to listen to some demos there (for me the generated quality is so so).
• Interactive Demo: users can interact with Bark model.

Coqui Github

• Interactive Demo
• Allows user to input voice and text so model can imitate your voice to speak out the texts.

[Speechify (OpenAI)]

• Can try API for free (limited hrs)
• Pricing information is listed here

Evaluations:

• Log F0
• Human eval in the loop
• mel cepstral distortion (MCD)

At the end, I wanna record some technical terms encountered when reading papers so as to easy recall them in the future:

• Speaker Verification
• (log-)mel Spectrogram