Minimal Example for English Text-to-Speech with VITS: Female and Male Voices

VITS (Variational Inference with adversarial learning for end-to-end Text-to-Speech) is a popular neural TTS architecture that combines text-to-speech and vocoding in a single model. It produces natural-sounding speech with relatively simple APIs.

In this post, we show a minimal example of generating:

• An English female voice (single-speaker)
• An English male voice (multi-speaker, speaker-selectable)

using open VITS models via the Coqui TTS library.

---

What this example demonstrates

• How to load pretrained VITS models from Hugging Face
• How single-speaker and multi-speaker TTS models differ
• How to select a specific speaker ID for multi-speaker synthesis
• How to generate WAV files in just a few lines of Python

---

Install dependencies

pip install TTS torch torchaudio

For best performance, a GPU is recommended but not required.

---

Example 1: English female voice (single-speaker VITS)

This model is trained on LJSpeech, a single-speaker English dataset. Because it’s single-speaker, no speaker ID is required.

from TTS.api import TTS

tts_female = TTS("tts_models/en/ljspeech/vits")

text = "Hello, I am an English female voice powered by VITS."

tts_female.tts_to_file(
    text=text,
    file_path="female.wav"
)

Output:

female.wav

---

Example 2: English male voice (multi-speaker VITS)

This model is trained on VCTK, a multi-speaker English dataset. You must specify a speaker ID to choose the voice.

from TTS.api import TTS

tts_male = TTS("tts_models/en/vctk/vits")

# Example speaker ID
speaker = "p252"

text = "Hello, I am an English male voice powered by VITS just for testing."

tts_male.tts_to_file(
    text=text,
    speaker=speaker,
    file_path="male.wav"
)

Output:

male.wav

---

About speaker IDs (VCTK)

The VCTK model includes many speakers, typically labeled:

p225, p226, p227, ... p300

In practice:

• Many male voices fall roughly in the range p225–p272
• Female voices appear throughout the remaining IDs

Exact gender or voice characteristics vary by speaker — experimentation is encouraged.

You can list available speakers programmatically:

print(tts_male.speakers)

---

When to use each model

---

Practical tips

• Single-speaker models are simpler and more consistent

• Multi-speaker models give flexibility but require speaker selection

• Audio quality improves with:

  • Shorter sentences
  • Proper punctuation
  • GPU inference (when available)

---

Limitations

• VITS models are not optimized for real-time synthesis
• Voice emotion and style control are limited
• Speaker identity is fixed by the dataset (no voice cloning)

For voice cloning or multilingual output, consider XTTS v2 instead.

---

Final thoughts

VITS remains a strong baseline for high-quality, open-source English TTS. With just a few lines of code, you can generate natural-sounding speech and explore different speaker identities.

If you’re building or learning TTS systems, this is an excellent place to start.

---

Are you using single-speaker TTS for consistency, or multi-speaker models for flexibility?