manim-voiceover-qwen3-tts

High-quality text-to-speech for Manim animations using Qwen3-TTS

A manim-voiceover plugin that integrates Alibaba's state-of-the-art Qwen3-TTS models, bringing natural-sounding voiceovers to your mathematical animations.

---

Features

Feature	Description
Voice Cloning	Clone any voice from a 3+ second audio sample
Voice Design	Create custom voices from natural language descriptions
Preset Voices	9 premium built-in voices with emotion/style control
Multi-language	Support for 10 languages including English, Chinese, Japanese, Korean
Caching	Automatic audio caching for fast re-renders
Multiple Characters	Easy voice switching for dialogue scenes

---

Installation

Option 1: Add to Existing Manim Project

If you already have manim and manim-voiceover installed:

pip install manim-voiceover-qwen3-tts

Option 2: Fresh Install with UV (Linux)

Set up a complete environment from scratch using UV:

# Install UV if you don't have it
curl -LsSf https://astral.sh/uv/install.sh | sh

# Install system dependencies for manim (Ubuntu/Debian)
sudo apt-get install -y libcairo2-dev libpango1.0-dev ffmpeg libsndfile1

# Create project directory
mkdir my-manim-project && cd my-manim-project

# Initialize UV project
uv init

# Add numba first (ensures correct dependency resolution)
uv add numba

# Add manim packages
uv add manim manim-voiceover manim-voiceover-qwen3-tts

# Copy the Quick Start example from the README (Option 1: Preset Voices section below)
# and save it as scene.py, then run:
uv run manim -pql scene.py QuickStart

Option 3: From Source

git clone https://github.com/DurhamSmith/manim-voiceover-qwen3-tts.git
cd manim-voiceover-qwen3-tts
pip install -e .

Optional: FlashAttention 2

For faster inference (requires compatible GPU):

pip install flash-attn --no-build-isolation

Requirements

• Python 3.10+
• CUDA-capable GPU (recommended, ~4GB VRAM for 1.7B models)
• manim >= 0.18.0
• manim-voiceover >= 0.3.0

System Dependencies

Manim requires some system libraries. On Ubuntu/Debian:

sudo apt-get install -y libcairo2-dev libpango1.0-dev ffmpeg libsndfile1

On macOS:

brew install cairo pango ffmpeg libsndfile

---

Quick Start

Option 1: Preset Voices (Easiest)

Use Qwen3's built-in premium voices - no setup required:

from manim import *
from manim_voiceover import VoiceoverScene
from manim_voiceover_qwen3_tts import Qwen3PresetVoiceService

class QuickStart(VoiceoverScene):
    def construct(self):
        self.set_speech_service(
            Qwen3PresetVoiceService(
                speaker="Ryan",
                language="English",
                use_flash_attention=False,  # Set True if flash-attn installed
            )
        )

        circle = Circle(color=BLUE)
        with self.voiceover(text="Let's draw a circle!") as tracker:
            self.play(Create(circle), run_time=tracker.duration)

Available Preset Speakers:

Language	Speakers
English	Ryan, Aiden
Chinese	Vivian, Serena, Uncle_Fu, Dylan, Eric
Japanese	Ono_Anna
Korean	Sohee

Option 2: Voice Design

Create any voice by describing it in natural language:

from manim import *
from manim_voiceover import VoiceoverScene
from manim_voiceover_qwen3_tts import Qwen3VoiceDesignService

class VoiceDesignDemo(VoiceoverScene):
    def construct(self):
        self.set_speech_service(
            Qwen3VoiceDesignService(
                voice_description="A warm, friendly female voice with a slight "
                                  "British accent, speaking clearly and professionally.",
                language="English",
                use_flash_attention=False,  # Set True if flash-attn installed
            )
        )

        title = Text("Welcome!")
        with self.voiceover(text="Welcome to our tutorial!") as tracker:
            self.play(Write(title), run_time=tracker.duration)

Option 3: Voice Cloning

Clone any voice from a short audio sample (3+ seconds):

from manim import *
from manim_voiceover import VoiceoverScene
from manim_voiceover_qwen3_tts import Qwen3VoiceCloningService, VoiceProfile

# Define a voice profile
narrator = VoiceProfile(
    name="narrator",
    ref_audio="voices/narrator_sample.wav",  # Your audio file
    ref_text="This is a sample of the narrator speaking clearly.",  # Transcript
    language="English",
)

class VoiceCloneDemo(VoiceoverScene):
    def construct(self):
        self.set_speech_service(
            Qwen3VoiceCloningService(
                voices=[narrator],
                default_voice="narrator",
                use_flash_attention=False,  # Set True if flash-attn installed
            )
        )

        with self.voiceover(text="Hello! My voice was cloned from a short sample.") as tracker:
            self.wait(tracker.duration)

---

Multi-Character Dialogue

Perfect for educational videos with multiple speakers:

from manim import *
from manim_voiceover import VoiceoverScene
from manim_voiceover_qwen3_tts import Qwen3VoiceCloningService, VoiceProfile

# Define character voices
alice = VoiceProfile(
    name="alice",
    ref_audio="voices/alice.wav",
    ref_text="Hi, I'm Alice and I love explaining math concepts!",
)

bob = VoiceProfile(
    name="bob",
    ref_audio="voices/bob.wav",
    ref_text="Hey there, I'm Bob. Let me ask you a question.",
)

class DialogueScene(VoiceoverScene):
    def construct(self):
        self.set_speech_service(Qwen3VoiceCloningService(
            voices=[alice, bob],
            use_flash_attention=False,  # Set True if flash-attn installed
        ))

        # Visual setup
        alice_label = Text("Alice", color=BLUE).to_edge(LEFT)
        bob_label = Text("Bob", color=RED).to_edge(RIGHT)
        self.add(alice_label, bob_label)

        # Dialogue
        with self.voiceover(text="Hi Bob! Want to learn about vectors?", voice="alice"):
            self.play(Indicate(alice_label))

        with self.voiceover(text="Sure Alice! That sounds interesting.", voice="bob"):
            self.play(Indicate(bob_label))

        with self.voiceover(text="Great! A vector has both magnitude and direction.", voice="alice"):
            arrow = Arrow(LEFT, RIGHT, color=YELLOW)
            self.play(Create(arrow))

---

API Reference

Services

Qwen3PresetVoiceService

Use Qwen3's premium preset voices with optional emotion/style control.

Qwen3PresetVoiceService(
    speaker="Ryan",                                      # Preset speaker name
    language="English",                                  # Language for synthesis
    instruct="Speak with enthusiasm",                    # Optional: style instruction
    model="Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice",       # Model ID
    device="cuda:0",                                     # Device (cuda:0, cpu)
    dtype="bfloat16",                                    # Weight dtype
    use_flash_attention=True,                            # Use FlashAttention 2
    output_format="mp3",                                 # Output format (mp3/wav)
)

Qwen3VoiceDesignService

Create custom voices from natural language descriptions.

Qwen3VoiceDesignService(
    voice_description="Description of desired voice characteristics",
    language="English",
    model="Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign",
    device="cuda:0",
    dtype="bfloat16",
    use_flash_attention=True,
    output_format="mp3",
)

Qwen3VoiceCloningService

Clone voices from reference audio samples.

Qwen3VoiceCloningService(
    voices=[voice_profile1, voice_profile2],            # List of VoiceProfile objects
    default_voice="narrator",                            # Default voice name
    model="Qwen/Qwen3-TTS-12Hz-1.7B-Base",
    device="cuda:0",
    dtype="bfloat16",
    use_flash_attention=True,
    output_format="mp3",
)

Classes

VoiceProfile

Define a voice for cloning.

VoiceProfile(
    name="character_name",           # Unique identifier for this voice
    ref_audio="path/to/audio.wav",   # Reference audio file (3+ seconds)
    ref_text="Transcript of audio",  # Exact transcript of the reference audio
    language="Auto",                 # Language ("Auto" for auto-detection)
)

Per-Voiceover Overrides

Override any setting for individual voiceover calls:

# Override speaker
with self.voiceover(text="Hello!", speaker="Aiden") as tracker:
    ...

# Override voice (for cloning service)
with self.voiceover(text="Hello!", voice="bob") as tracker:
    ...

# Override style instruction
with self.voiceover(text="Wow!", instruct="Speak with excitement") as tracker:
    ...

# Override language
with self.voiceover(text="Bonjour!", language="French") as tracker:
    ...

---

Available Models

Model	Parameters	Use Case	VRAM
Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice	1.7B	Preset voices	~4GB
Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign	1.7B	Voice design	~4GB
Qwen/Qwen3-TTS-12Hz-1.7B-Base	1.7B	Voice cloning	~4GB
Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice	0.6B	Lightweight preset	~2GB
Qwen/Qwen3-TTS-12Hz-0.6B-Base	0.6B	Lightweight cloning	~2GB

---

Supported Languages

All services support 10 languages:

• English
• Chinese (Mandarin)
• Japanese
• Korean
• German
• French
• Russian
• Portuguese
• Spanish
• Italian

---

Performance Tips

1. Enable FlashAttention 2

Significantly faster inference on compatible GPUs:

pip install flash-attn --no-build-isolation

2. Use Smaller Models

For faster generation with acceptable quality:

Qwen3PresetVoiceService(
    model="Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice",
    ...
)

3. Leverage Caching

manim-voiceover automatically caches generated audio. Re-renders with unchanged text are instant.

4. Voice Prompt Caching

For voice cloning, the service automatically caches voice prompts. The first generation with a new voice takes longer, but subsequent uses are fast.

---

Troubleshooting

CUDA Out of Memory

Option 1: Use a smaller model:

Qwen3PresetVoiceService(
    model="Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice",
)

Option 2: Run on CPU (slower):

Qwen3PresetVoiceService(
    device="cpu",
    dtype="float32",
    use_flash_attention=False,
)

FlashAttention Not Available

Disable it explicitly:

Qwen3PresetVoiceService(
    use_flash_attention=False,
)

Audio Quality Issues

• Ensure reference audio is at least 3 seconds long for voice cloning
• Use high-quality reference audio (clear speech, minimal background noise)
• Verify the transcript exactly matches the reference audio

Model Download Issues

Models are downloaded from HuggingFace on first use. Ensure you have:

• Stable internet connection
• Sufficient disk space (~7GB for 1.7B models)

Known Warnings

You may see deprecation warnings when running. These come from upstream dependencies, not this package:

UserWarning: pkg_resources is deprecated as an API...
FutureWarning: librosa.core.audio.__audioread_load Deprecated...
UserWarning: PySoundFile failed. Trying audioread instead.

Warning	Source	Status
pkg_resources deprecated	manim-voiceover	Upstream issue - awaiting fix
librosa.__audioread_load deprecated	qwen-tts → librosa	Upstream issue - awaiting fix
PySoundFile failed	qwen-tts → librosa	Install libsndfile (see below)

To reduce warnings:

1. Install system audio library:

   # Ubuntu/Debian
   sudo apt-get install libsndfile1
   
   # macOS
   brew install libsndfile

2. Suppress warnings in your script (optional):

   import warnings
   warnings.filterwarnings("ignore", category=DeprecationWarning)
   warnings.filterwarnings("ignore", category=FutureWarning)

These warnings don't affect functionality - your videos will render correctly.

---

Voice Cloning Best Practices

Reference Audio Guidelines

1. Duration: 3-10 seconds is ideal
2. Quality: Clear audio without background noise
3. Content: Natural speech, not whispered or shouted
4. Format: WAV or MP3 supported

Transcript Accuracy

The transcript must exactly match what's said in the reference audio. This helps the model understand the voice characteristics.

Organizing Voice Profiles

For projects with multiple characters, organize your voices:

project/
├── voices/
│   ├── narrator/
│   │   ├── sample.wav
│   │   └── metadata.json
│   ├── teacher/
│   │   ├── sample.wav
│   │   └── metadata.json
│   └── student/
│       ├── sample.wav
│       └── metadata.json
├── scenes/
│   └── my_scene.py

---

Examples

See the examples/ directory for complete working examples:

Example	Service	Description
preset_voices.py	Qwen3PresetVoiceService	Preset speakers + languages — switches between built-in voices (Ryan, Vivian, Ono_Anna, Sohee) across English, Chinese, Japanese, Korean
emotion_showcase.py	Qwen3PresetVoiceService	One voice, many emotions — same speaker (Ryan), varying instruct per line (happy, sad, angry, excited, calm, etc.)
voice_design.py	Qwen3VoiceDesignService	Many designed voices — same content delivered by 4 different voices created from text descriptions
storytelling_scene.py	Qwen3VoiceDesignService	Multi-character story — narrator, hero, mentor, villain each with unique designed voices
voice_cloning.py	Qwen3VoiceCloningService	Clone from audio — clone voices from reference .wav files, switch between multiple cloned voices

Note: voice_cloning.py includes a sample narrator voice (voices/narrator.wav). To add your own voices, create additional VoiceProfile entries with:

• ref_audio: path to your .wav file (3+ seconds of clear speech)
• ref_text: exact transcript of what's spoken in the audio

Running Examples

# Preset voices (built-in speakers, multiple languages)
manim -pql examples/preset_voices.py PresetVoicesDemo

# Emotion control (same voice, different emotions via instruct)
manim -pql examples/emotion_showcase.py EmotionShowcase

# Voice design (create voices from descriptions)
manim -pql examples/voice_design.py VoiceDesignDemo

# Storytelling (multi-character with designed voices)
manim -pql examples/storytelling_scene.py StorytellingScene

# Voice cloning (requires your own .wav files)
manim -pql examples/voice_cloning.py VoiceCloningDemo

---

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

1. Fork the repository
2. Create your feature branch (git checkout -b feature/amazing-feature)
3. Commit your changes (git commit -m 'Add amazing feature')
4. Push to the branch (git push origin feature/amazing-feature)
5. Open a Pull Request

---

License

This project is licensed under the MIT License - see the LICENSE file for details.

---

Acknowledgments

• Qwen3-TTS - The underlying TTS model by Alibaba
• manim-voiceover - The voiceover framework this plugin extends
• Manim Community - The amazing animation library

---

Citation

If you use this project in your research or videos, please consider citing:

@software{manim_voiceover_qwen3_tts,
  title = {manim-voiceover-qwen3-tts: Qwen3-TTS Integration for Manim},
  url = {https://github.com/DurhamSmith/manim-voiceover-qwen3-tts},
  year = {2026}
}