On-Device Text-to-Speech Middleware
No cloud required. No heavy dependencies like PyTorch.
Just pip install — and reproduce any voice from a 10-second reference clip.
WHY ailia AI Voice
A third option — different from cloud API-based TTS and from running open-source software as-is.
No PyTorch required — it runs on ONNX, so there's no framework version hell.
In Python, it's a single pip install ailia_voice command.
Model files download automatically, and speech synthesis is complete in just a few lines of code.
Powered by GPT-SoVITS — just provide about 10 seconds of reference audio, and synthesize speech in that voice and speaking style. No training or fine-tuning required. Switch freely between voices, from game characters to narration.
ailia provides its own distilled high-speed model. Compared to the open-source GPT-SoVITS v2-pro, it achieves 1.8x faster inference*, enabling near real-time response even on GPU-less edge devices and mobile. You can switch to the standard model depending on your use case.
* Based on our internal measurements.
DEMO
From a single Japanese reference clip (about 10 seconds), the same voice reads aloud in Japanese, English, and Chinese.
No training or fine-tuning was performed — every result below is a zero-shot synthesis.
INPUT — Reference audio
Using just one Japanese reference audio file as input, all three languages below are synthesized.
Japanese
「こんにちは。今日はいい天気ですね。」
English
"Hello. It's a beautiful day today."
Chinese
「你好。今天天气真好。」
ailia's own distilled model is 1.8x faster* than the open-source GPT-SoVITS v2-pro. Compare synthesis results using the same reference audio and same text as above.
* Based on our internal measurements.
This demo's speech synthesis uses voice data freely released by the free material character "Tsukuyomi-chan" (© Rei Yumesaki).
■Tsukuyomi-chan Corpus (CV: Rei Yumesaki)
USE CASES
Speech synthesis AI can be used in many different scenarios.
Add a voice to an LLM-powered chatbot to create a talking AI agent. Combine with ailia LLM and AI Speech for a fully offline conversational system.
Naturally reads translated text aloud in Japanese, English, and Chinese. Works even in locations without a network connection.
Synthesizes speech in another language while preserving the original voice quality. Also useful for multilingual voice guidance on industrial equipment.
Automatically generate narration for presentation videos and video storyboards from text.
With Unity support, in-game characters can speak in real time.
CASE STUDY
Art / Experiential Pavilion
In this experiential pavilion where visitors converse with their own AI avatar, "Mirrored Body®," ailia was used to implement speech recognition, speech synthesis, low-latency edge AI, and generative AI. It supported an interactive experience for hundreds of thousands of visitors, with the related app used by over 150,000 people.
ailia AI Voice has also been used to produce narration for concerts by the Japan Philharmonic Orchestra, directed by Mr. Ochiai.
Read the interview →MODELS
Switch models just by changing an argument — the API stays the same.
Use the high-quality model for prototyping and the distilled model for production, for example.
| Model | Features | Speed | Quality | Recommended Use |
|---|---|---|---|---|
| Tacotron2 | English baseline | Fast | Standard | Lightweight, English-only use |
| GPT-SoVITS v1 | Lightest-weight voice clone | Fast | Good | Resource-constrained environments |
| GPT-SoVITS v2 | Supports Japanese accent and speech rate control | Slightly fast | Good | Standard choice for real-time use |
| GPT-SoVITS v2-pro | v3 text analysis + v2 fast vocoder + speaker verification | Slightly fast | High quality | Balance of quality and speed (Recommended) |
| GPT-SoVITS v3 | Highest audio quality via diffusion | Medium speed | Best | Narration and content production |
| ailia's own distilled model | Our own distillation of GPT-SoVITS v2-pro. 1.8x faster than the OSS model* | Fastest | Good | Real-time response for mobile and embedded |
Each model is available in Japanese, English, and Chinese versions. Output is 32 kHz mono. * Based on our internal measurements.
VS CLOUD TTS
Cloud TTS is convenient, but for mass-produced products, "cost at scale" and "behavioral stability" become challenges.
ailia AI Voice charges only a license fee — no extra charges based on inference count or number of users.
| ailia AI Voice | Cloud TTS API | |
|---|---|---|
| API Cost |  No charges based on inference count or user count. License fee only — cost stays constant as usage grows |
 Pay-as-you-go pricing based on character count and requests. Costs scale with user count, making budgeting difficult at production scale |
| Behavioral Stability | The model doesn't auto-update, so behavior stays consistent — easy to validate and operate long-term for mass-produced products |
The provider may update the model, change the API spec, or discontinue the service, causing output to change without notice |
| Operational Stability | Runs entirely on-device. Unaffected by network or API outages, with no need for server monitoring |
Risk of downtime due to network outages, API failures, or rate limiting |
| Latency | No network round-trip. Fast response even on the edge with the distilled model |
 A network round-trip always occurs, depending on connection quality |
| Data Privacy | Text and audio are never sent externally. Usable even for confidential information or in offline environments |
The text to be read aloud must be sent to an external server |
| Bundling into Apps | Bundle it into your game or app build and run it on the end user's device. Ship your product with redistribution rights included |
Requires a cloud connection, making it unsuitable for offline products or environments without connectivity |
| Japanese Reading Control | Reliably control the reading of proper nouns and technical terms with a user dictionary (pyopenjtalk format). Also supports Japanese pitch accent |
Misreadings must be avoided through text-side workarounds, and reliable control options are often limited |
| Voice Cloning | Zero-shot with a 10-second reference clip. No additional cost |
Custom voices often require additional fees, review, and a training period |
* This reflects general tendencies of typical cloud-based speech synthesis services.
FEATURES
A rich feature set to support production use.
No Cloud
Runs entirely on edge devices. Since text is never sent to a server, confidential information stays safe. No pay-as-you-go API charges either.
Multilingual
Built-in OpenJTalk for Japanese phoneme conversion and g2pw + jieba for Chinese. Everything, including G2P, is self-contained in the library.
Customization
Load user dictionaries in pyopenjtalk format. Precisely control the pronunciation of proper nouns and technical terms.
Cross-Platform
Supports Windows / macOS / Linux / iOS / Android. Deploy to every platform with a single API.
Multi-Language Bindings
Prototype in Python, then deploy to mobile as-is with C++ / Unity (C#) / Flutter / Kotlin.
Stable Operation
Few dependencies, and the model never auto-updates, so behavior stays consistent — easy to validate for mass-produced products.
Control
Supports adjusting playback speed (v2 and later). Achieve easy-to-understand narration tailored to your use case.
GPU
Thanks to ailia SDK's CUDA implementation, you can leverage GPUs.
Integration
ailia LLM and AI Speech can be combined to build a conversational agent that listens, thinks, and speaks.
QUICK START
Models are downloaded automatically on first run. No more wrestling with requirements.
# Install (this is the only dependency)
pip3 install ailia_voice librosa soundfile
# Synthesize speech
import ailia_voice, librosa, soundfile
ref_audio, rate = librosa.load("reference_10sec.wav", mono=True)
ref_text = "This is the transcript of the reference audio."
voice = ailia_voice.GPTSoVITSV2Pro()
voice.initialize_model(model_path="./models/") # auto-downloads the model
voice.set_reference_audio(ref_text,
ailia_voice.AILIA_VOICE_G2P_TYPE_GPT_SOVITS_JA, ref_audio, rate)
buf, sr = voice.synthesize_voice("こんにちは。今日はいい天気ですね。",
ailia_voice.AILIA_VOICE_G2P_TYPE_GPT_SOVITS_JA)
soundfile.write("output.wav", buf, sr)PRICING
Evaluate for free first. Sign a license only once you decide to go into production.
There are never extra charges based on inference count.
TIER 01
Free· No credit card required
All features available for 30 daysSupports all models (every GPT-SoVITS version / the distilled model)Includes all platformsFor internal testing onlyTIER 02
Annual subscription(per app) — device-based pricing also available
Embed in production environments and commercial appsRedistribution rights to end usersNo extra charges based on inference count or user countEmail and meeting supportSERVICE
Custom quote
Custom distilled models and optimization for specific devicesSupport implementing voice features in your app (ailia WORKS)End-to-end support from consulting to model creationOnce you've confirmed it works with the evaluation version, let us know your use case, device count, and app configuration. We'll propose the best license for you. Consultations and quotes are free.
SUPPORT
We'll support your development from adoption through production.
Covers setup instructions, API details, and platform-specific references.
We publish articles on API usage examples and model explanations.
Ask developers questions directly, chat-based.
Try every feature for free with a one-month evaluation license.
Feel free to reach out about license types, target devices, custom distilled models, and more.
Companies using the ailia SDK
