Compare voices
Compare the speech engines available in Mutelink and listen to voices used by their standard behavior before downloading.
Engine differences
| VOICEVOX | Irodori-TTS | Qwen3-TTS | |
|---|---|---|---|
| Languages | Japanese | Japanese | Japanese, English, Chinese, Korean, German, Spanish, French, Italian, Portuguese, Russian |
| Requirements | No GPU | Vulkan-compatible GPU | NVIDIA GPU and acceleration recommended for conversation |
| Download | ~30MB engine plus character files | ~1.9GB | ~1.2GB standard model + ~395MB NVIDIA acceleration; cloning adds ~1.6GB |
| Voice | Select a character | Standard or registered voice | Automatically selected standard voice for each language, or registered voice |
| Free features | All | Standard voices | Standard voices |
| Best for | No GPU or character voices | Fast, natural Japanese | Speech in languages beyond Japanese |
- Voice registration (cloning) for both Irodori-TTS and Qwen3-TTS requires the supporter edition. Qwen cloning needs an additional ~1.6GB model.
- Add Qwen3-TTS NVIDIA acceleration (~395MB) in Settings > Qwen3-TTS.
- Irodori-TTS and Qwen3-TTS share VRAM with VRChat. Memory use varies with text length.
- Each VOICEVOX character has its own terms. Check before downloading.
Before using Qwen3-TTS for conversation
The developer measured RTF above 2 without NVIDIA acceleration (Vulkan): generation could not keep up with playback. RTF 2 means 10 seconds to generate 5 seconds of speech. With other GPU workloads, RTF was ~3.0–3.3.
For real-time conversation, plan on an NVIDIA GPU and acceleration. Cloning needs both the cloning model and NVIDIA acceleration. Before buying, test free standard voices with acceleration. Speed also depends on your PC and VRChat load.
Listen to samples
Registered voices are on the reference/cloning comparison page: Japanese Irodori and Qwen’s four references × four output languages.
Examples of standard Irodori-TTS and Qwen3-TTS voices.
About Qwen3-TTS standard voices
By default, a voice is selected automatically from candidates for the language each time speech is synthesized. These samples are examples.
Japanese and Korean each have one candidate, so their speaker stays the same. Chinese has five; English and the other supported languages use two English speakers. Delivery can vary even with the same speaker.
Japanese (Irodori-TTS and Qwen3-TTS)
| Text | Irodori-TTS(standard) | Qwen3-TTS(standard) |
|---|---|---|
| こんにちは、はじめまして。 | ||
| すみません、聞き取れませんでした。もう一度言ってもらえますか? | ||
| 今からあなたのインスタンスに入りますね。 | ||
| そのアバター大好きです。どこで手に入れたんですか? |
Examples of different Qwen3-TTS voices
English and Chinese examples of the same sentence with different standard voices. Two of five Chinese candidates are included. These are saved recordings; replaying them does not change the voice.
| Language | Text | Standard voice (example 1) | Standard voice (example 2) |
|---|---|---|---|
| English | Hello, nice to meet you. | ||
| Chinese | 你好,很高兴认识你。 |
Other languages (Qwen3-TTS)
| Language | Text | Qwen3-TTS(standard) |
|---|---|---|
| English | Hello, nice to meet you. | |
| English | I'm sorry, I didn't catch that. Could you say it again? | |
| Chinese | 你好,很高兴认识你。 | |
| Chinese | 对不起,我没听清楚。你能再说一遍吗? | |
| Korean | 안녕하세요, 만나서 반갑습니다. | |
| Korean | 죄송해요, 잘 못 들었어요. 다시 말씀해 주시겠어요? | |
| German | Hallo, schön, dich kennenzulernen. | |
| German | Entschuldigung, das habe ich nicht verstanden. Könntest du es bitte wiederholen? | |
| Spanish | Hola, mucho gusto. | |
| Spanish | Lo siento, no te he entendido. ¿Podrías repetirlo? | |
| French | Bonjour, enchanté de faire votre connaissance. | |
| French | Désolé, je n'ai pas compris. Pourrais-tu répéter ? | |
| Italian | Ciao, piacere di conoscerti. | |
| Italian | Scusa, non ho capito. Potresti ripetere? | |
| Portuguese | Olá, prazer em conhecer você. | |
| Portuguese | Desculpe, não entendi. Você poderia repetir? | |
| Russian | Здравствуйте, рад с вами познакомиться. | |
| Russian | Извини, я не расслышал. Можешь повторить? |
Sample and measurement conditions
Samples fix the standard voice candidate, normally generating once with seed 41001 before MP3 conversion. Example 1/2 distinguish recordings, not selectable app speakers. Normal app synthesis changes the seed on every request.
- The Russian Qwen3-TTS sample “Здравствуйте, рад с вами познакомиться.” was regenerated with seed 769694441. Speed/error figures remain the original 30-sentence measurements.
- The Japanese Irodori-TTS sample “こんにちは、はじめまして。” was regenerated with seed 10990944. Speed/error figures remain the original 30-sentence measurements.
- The Japanese Irodori-TTS sample “今からあなたのインスタンスに入りますね。” was regenerated with seed 1362122378. Speed/error figures remain the original 30-sentence measurements.
| Sample | Voice used |
|---|---|
| Irodori-TTS(Japanese) | No reference voice(standard) |
| Qwen3-TTS(English) | ryan (example 1)、aiden (example 2) |
| Qwen3-TTS(Chinese) | vivian (example 1)、serena (example 2) |
| Qwen3-TTS(Korean) | sohee |
| Qwen3-TTS(German) | ryan |
| Qwen3-TTS(Spanish) | ryan |
| Qwen3-TTS(French) | ryan |
| Qwen3-TTS(Italian) | ryan |
| Qwen3-TTS(Portuguese) | ryan |
| Qwen3-TTS(Russian) | ryan |
| Qwen3-TTS(Japanese) | ono_anna |
Speed/error rates used example-1 speakers fixed per language (ono_anna for Japanese; sohee for Korean), 30 sentences each. Additional example-2 samples are excluded from those aggregates.
VOICEVOX voices differ by character. Listen on the official VOICEVOX site.
Synthesis speed
Time needed to generate one sentence. Time/audio is generation time divided by audio duration: lower is faster; below 1 means generation can finish while playback proceeds.
| Engine | Language | Mean audio duration | Mean generation time | Time/audio |
|---|---|---|---|---|
| Irodori-TTS | Japanese | 3.8s | 0.7s | 0.19 |
| Qwen3-TTS | Japanese | 5.1s | 3.4s | 0.67 |
| Qwen3-TTS | English | 3.9s | 2.6s | 0.65 |
| Qwen3-TTS | Chinese | 3.2s | 2.0s | 0.63 |
| Qwen3-TTS | Korean | 4.5s | 2.6s | 0.58 |
| Qwen3-TTS | German | 3.3s | 1.1s | 0.33 |
| Qwen3-TTS | Spanish | 3.8s | 1.6s | 0.41 |
| Qwen3-TTS | French | 3.3s | 1.1s | 0.32 |
| Qwen3-TTS | Italian | 3.4s | 1.3s | 0.39 |
| Qwen3-TTS | Portuguese | 3.8s | 1.6s | 0.42 |
| Qwen3-TTS | Russian | 4.6s | 1.8s | 0.40 |
Indicative results from 30 sentences each on the developer’s RTX 4070 with NVIDIA acceleration. Qwen used one fixed standard speaker per language, not an average across candidates. Speaker details are in the sample conditions above.
Playback starts as chunks become available, without waiting for the full sentence.
Intelligibility reference
Synthesized audio was transcribed and compared with its source text. Values are mean recognition error rates across engines; lower means closer recognition. Japanese is compared by pronunciation.
| Engine | Language | Mean error rate |
|---|---|---|
| Irodori-TTS | Japanese | 7.7% |
| Qwen3-TTS | Japanese | 13.2% |
| Qwen3-TTS | English | 1.9% |
| Qwen3-TTS | Chinese | 2.8% |
| Qwen3-TTS | Korean | 3.2% |
| Qwen3-TTS | German | 4.3% |
| Qwen3-TTS | Spanish | 3.6% |
| Qwen3-TTS | French | 5.1% |
| Qwen3-TTS | Italian | 7.9% |
| Qwen3-TTS | Portuguese | 6.7% |
| Qwen3-TTS | Russian | 5.5% |
Limitations of these figures
- Qwen results use the same fixed speakers as the speed test. They do not represent average performance under automatic voice selection.
- Each language uses 30 short everyday sentences, AI-authored without native-speaker review.
- Recognition quirks such as number formatting and paraphrases also count as errors. These numbers do not measure voice quality alone.
- These figures do not show naturalness or personal preference. Listen to the samples above.
See the per-engine breakdown on speech recognition comparisons.