Clonación de voz — TTS Pruebak

🎤 Comparativa: Clonación de voz

Voz clonada desde el audio de Lazkao Txiki (EUS) usando 5 modelos diferentes · Samples en INGLÉS y Euskera · Cross-lingual voice cloning

🔷 REFERENCIA ORIGINAL

Lazkao Txiki (original) Referencia

0.0s 5168 KB Euskera · Bertsos ⬇

🔶 F5-TTS (flow-matching, ~24x RTF)

F5-TTS · Texto general 19.3s

19.3s 904 KB 24kHz ⬇

Texto

"Welcome to this test. This is a voice cloned from Lazkao Txiki using F5 TTS. The quick brown fox jumps over the lazy dog."

F5-TTS · Texto financiero 25.6s

25.6s 1198 KB 24kHz ⬇

Texto

"Numbers are handled naturally here. Five point two million dollars and forty two percent. This is how this voice sounds when reading financial data in English."

🔶 MOSS-TTS-Nano (Audio Tokenizer + LLM, 0.1B)

Nano · Texto general 100M

0.0s 2910 KB 48kHz est. ⬇

Texto

"Welcome to this test. This is a voice cloned from Lazkao Txiki using MOSS TTS Nano. The quick brown fox jumps over the lazy dog."

Nano · Texto financiero 100M

0.0s 3300 KB 48kHz est. ⬇

Texto

"Numbers are handled naturally here. Five point two million dollars and forty two percent. This is how this voice sounds when reading financial data in English."

🔶 MOSS-TTS-Realtime (1.7B/2.33B, LLM + Codec)

Realtime · Texto general 2.33B

0.0s 1530 KB 48kHz estéreo ⬇

Texto

"Welcome to this test. This is a voice cloned from Lazkao Txiki using MOSS TTS Realtime model. The quick brown fox jumps over the lazy dog."

Realtime · Texto financiero 2.33B

0.0s 1530 KB 48kHz estéreo ⬇

Texto

"Numbers are handled naturally here. Five point two million dollars and forty two percent. This is how this voice sounds when reading financial data in English."

🔶 OmniVoice GGUF (0.6B, MaskGIT, C++ GGML) — Textos en INGLÉS

OmniVoice · Scene 1: Investor Paradox 0.6B · ~26x RTF

30.4s 1427 KB 24kHz · Cross-lingual (ref: EU → out: EN) ⬇

Texto

"Most investors do everything right. They research their assets, they diversify, they stay patient, and they still underperform the market by two to four percent every single year. Not because of bad picks."

🔶 Qwen3-TTS GGUF (0.6B, AR, C++ GGML) — Textos en INGLÉS 🚀

Qwen3-TTS · Scene 1: Investor Paradox 0.6B · ~1.8x RTF

15.4s 724 KB 24kHz · Cross-lingual (ref: EU → out: EN) ⬇

Texto

Qwen3-TTS · QuantGuard Intro (texto largo) 0.6B · ~3.4x RTF

21.8s 1020 KB 24kHz · 21.8s en 74s de proceso ⬇

Texto

"Most investors do everything right. They research their assets. They diversify. They stay patient through the ups and downs. And they still underperform the market by two to four percent every single year. Not because of bad picks. Because of three mistakes they did not even know they were making. Mistakes that a simple algorithm can eliminate completely."

🔶 Qwen3-TTS GGUF 1.7B (AR, C++ GGML) — Textos en INGLÉS 🚀🏆

Qwen3-TTS 1.7B · Scene 1: Investor Paradox 1.7B · ~2.4x RTF

13.0s 611 KB 24kHz · Cross-lingual (ref: EU → out: EN) ⬇

Texto

Qwen3-TTS 1.7B · QuantGuard Intro (texto largo) 1.7B · ~2.4x RTF

19.5s 915 KB 24kHz · Raw TTS quality ⬇

Texto

🔶 OuteTTS 1.0-0.6B (LLM + DAC codec)

OuteTTS · Euskera (original) 0.6B

2.7s 237 KB 44.1kHz ⬇

Texto

"Welcome to this test. This is a voice cloned from Lazkao Txiki using OuteTTS."

🗣️ Samples en euskera (originales)

OmniVoice · Euskera (13.7s)

Qwen3-TTS 0.6B · Euskera (12.2s)

Qwen3-TTS · Euskera (alternativo)

📊 Comparativa de modelos

Modelo	Params	Sample rate	Canales	RTF (CPU)	Clonación	Idiomas
F5-TTS	~335M	24 kHz	Mono	~24x	✅ Zero-shot	Cualquiera
MOSS-TTS-Nano	100M	48 kHz	Estéreo	~8x	✅ Zero-shot	20
MOSS-TTS-Realtime	2.33B	48 kHz	Estéreo	~15x	✅ Zero-shot	20
OmniVoice GGUF	0.6B	24 kHz	Mono	~26x	✅ Zero-shot	646
Qwen3-TTS GGUF 0.6B	0.6B	24 kHz	Mono	~2.2x	✅ Zero-shot	11
Qwen3-TTS GGUF 1.7B	1.7B	24 kHz	Mono	~2.4x	✅ Zero-shot	11
OuteTTS 1.0	0.6B	44.1 kHz	Mono	~?	✅ Zero-shot	En