Reference audio and voice cloning
Hear how Irodori-TTS and Qwen3-TTS reproduce a registered reference voice. Both the reference and generated audio can be played.
The original voice is VOICEVOX:小夜/SAYO. A Japanese WAV exported from Sayo was used on an external cloning site to create Japanese, English, Chinese and Korean references. Those four references were then registered with Irodori and Qwen.
Compare the same sentence using a Japanese reference versus a reference in the output language. Qwen has all 4 reference × 4 output combinations (16); Irodori has Japanese → Japanese (1).
Voice cloning requires the supporter edition. See speech engine comparison for Qwen models and acceleration needed for conversation.
Original Sayo audio
The source is a Japanese WAV exported from VOICEVOX:小夜/SAYO (Normal). It was used on an external cloning site to create the following four-language references.
Transcript of the original Sayo audio
一緒に冒険しよう、絶対楽しいことが待ってるよ。えーっと、これでいいのかな、ちょっと心配だけど、やってみるね。ねえねえ、聞いて、ちょっとだけ秘密の話を教えてあげる。
Reference recordings
All four references were generated externally from the Sayo WAV above, including the Japanese reference. Play them to compare voice similarity with the generated samples.
| Reference language | Audio | Duration |
|---|---|---|
| Japanese | 6.60s | |
| Chinese | 7.37s | |
| Korean | 6.70s | |
| English | 4.61s |
Reference transcripts and provenance
Japanese: generated on an external site from the Sayo reference WAV
そのアバター大好きです。どこで手に入れたんですか?私も使ってみたいです。
Chinese: generated on an external site from the Sayo reference WAV
我很喜欢那个虚拟形象。请问是在哪里弄到的?我也想用用看。
Korean: generated on an external site from the Sayo reference WAV
그 아바타 정말 마음에 드네요. 어디서 구하셨나요? 저도 한번 써보고 싶어요.
English: generated on an external site from the Sayo reference WAV
I love that avatar! Where did you get it? I’d like to use it too.
Original voice: VOICEVOX:小夜/SAYO. These samples register the externally generated Sayo-derived references with Irodori-TTS or Qwen3-TTS, then synthesize speech.
Japanese:Irodori-TTS
Text: こんにちは。今日はどんな一日でしたか?
Generated using the registered Japanese reference.
Qwen3-TTS: 4 references × 4 output languages
All samples in the same output language use the same text and random seed. Matching reference languages are highlighted.
Speaking Japanese
Text: こんにちは。今日はどんな一日でしたか?
| Reference language | Qwen3-TTS output |
|---|---|
| Japanese(same language) | |
| Chinese | |
| Korean | |
| English |
Speaking Chinese
Text: 你好。你今天过得怎么样?
| Reference language | Qwen3-TTS output |
|---|---|
| Japanese | |
| Chinese(same language) | |
| Korean | |
| English |
Speaking Korean
Text: 안녕하세요. 오늘 하루는 어땠나요?
| Reference language | Qwen3-TTS output |
|---|---|
| Japanese | |
| Chinese | |
| Korean(same language) | |
| English |
Speaking English
Text: Hello. How was your day today?
| Reference language | Qwen3-TTS output |
|---|---|
| Japanese | |
| Chinese | |
| Korean | |
| English(same language) |
Generation settings
Each condition was generated once with seed 41001. Irodori-TTS used 4 steps and the Japanese reference; Qwen3-TTS used 0.6B Base Q8_0 with NVIDIA CUDA. Registration/synthesis servers are the same as the app. Audio was converted to mono 64kbps MP3.
Reference texts mean the same thing, but lengths and pronunciation differ, so differences cannot be attributed to language alone. Compare intelligibility, pronunciation, similarity and noise separately. These samples do not change the existing speed/error-rate aggregates.
Original voice: VOICEVOX:小夜/SAYO (Normal). Official Sayo site.