CosyVoice 3: Towards In-the-wild Speech Generation via Scaling-up and Post-training

Benchmark Model Rank Results
text-to-speech-synthesis-on-seed-tts-test-hardCosyVoice 3-0.5B RL#1CER: 5.09Speaker Similarity: 0.75
text-to-speech-synthesis-on-seed-tts-test-hardQwen2.5-Omni-7B RL#5CER: 6.54Speaker Similarity: 0.752
text-to-speech-synthesis-on-seed-tts-test-hardSeed-TTS#7CER: 7.59Speaker Similarity: 0.776
text-to-speech-synthesis-on-seed-tts-test-hardQwen2.5-Omni-7B#8CER: 7.97Speaker Similarity: 0.747
text-to-speech-synthesis-on-seed-tts-test-hardF5-TTS (32 NFE)#9CER: 8.67Speaker Similarity: 0.713
text-to-speech-synthesis-on-seed-tts-test-hardF5R-TTS#10CER: 8.79Speaker Similarity: 0.718
text-to-speech-synthesis-on-seed-tts-test-hardMaskGCT#11CER: 10.27Speaker Similarity: 0.748
text-to-speech-synthesis-on-seed-tts-test-hardCosyVoice#12CER: 11.75Speaker Similarity: 0.709
text-to-speech-synthesis-on-seed-tts-test-hardFireRedTTS#13CER: 17.45Speaker Similarity: 0.621
text-to-speech-synthesis-on-seed-tts-test-enCosyVoice 3-1.5B RL#6WER: 1.45
text-to-speech-synthesis-on-seed-tts-test-enE2 TTS (32 NFE)#13WER: 2.19Speaker Similarity: 0.71
text-to-speech-synthesis-on-seed-tts-test-enQwen2.5-Omni-7B RL#15WER: 2.33Speaker Similarity: 0.641
text-to-speech-synthesis-on-seed-tts-test-enQwen2.5-Omni-7B#19WER: 2.72Speaker Similarity: 0.632
text-to-speech-synthesis-on-seed-tts-test-enFireRedTTS#21WER: 3.82Speaker Similarity: 0.46
text-to-speech-synthesis-on-seed-tts-test-zhF5R-TTS#12CER: 1.37Speaker Similarity: 0.754
text-to-speech-synthesis-on-seed-tts-test-zhQwen2.5-Omni-7B RL#14CER: 1.42Speaker Similarity: 0.754
text-to-speech-synthesis-on-seed-tts-test-zhFireRedTTS#16CER: 1.51Speaker Similarity: 0.635
text-to-speech-synthesis-on-seed-tts-test-zhQwen2.5-Omni-7B#18CER: 1.7Speaker Similarity: 0.752
text-to-speech-synthesis-on-seed-tts-test-zhE2 TTS (32 NFE)#19CER: 1.97Speaker Similarity: 0.73
text-to-speech-synthesis-on-seed-tts-test-zhCosyVoice 3-1.5B RL#23Speaker Similarity: 0.775