FastSpeech 2 - Multilingual

运行方法

python synthesize.py --restore_step 900000 --mode mixed --text "Numbers如何寻找最优特征？是有放回还是无放回的呢？"  -p config/hcsi_10speakers/preprocess.yaml -m config/hcsi_10speakers/model.yaml -t config/hcsi_10speakers/train.yaml --speaker_id 5 --duration_control 1 --pitch_control 1 --energy_control 1

说话人

speaker 0-9：

• 0 Spk4.CN.03FR00

• 1 Spk9.EN.F.DB2

• 2 Spk8.EN.F.DB1

• 3 Spk2.CN.Deng

• 4 Spk1.CN.DataBaker

• 5 Spk6.CN.Pachira

• 6 Spk7.EN.M.DB1

• 7 Spk3.EN.XuYue

• 8 Spk5.CN.03MR00

• 9 Spk0.EN.LJSpeech

mfa提取

mfa align raw_data/hcsi_10speakers/Spk0.EN.LJSpeech/ lexicon/librispeech-lexicon.txt english preprocessed_data/hcsi_10speakers/Spk0.EN.LJSpeech/ --clean --disable_textgrid_cleanup --optional_silence_phone sp --other_noise_phone onp


# new_acoustic_model 放在 /ceph/home/huangqc18/Documents/MFA/pretrained_models/acoustic中
mfa adapt raw_data/hcsi_10speakers/Spk1.CN.DataBaker/ lexicon/pinyin-lexicon-r.txt new _acoustic_model preprocessed_data/hcsi_10speakers/Spk1.CN.DataBaker/ --clean --disable_textgrid_cleanup --optional_silence_phone sp --other_noise_phone onp

References

• FastSpeech 2: Fast and High-Quality End-to-End Text to Speech, Y. Ren, et al.
• ming024's FastSpeech implementation