Skip to content

SOTA Open Source TTS

Optimize sft config

Lengyuedecc7fe66e4d
diff --git a/fish_speech/configs/text2semantic_sft.yaml b/fish_speech/configs/text2semantic_sft.yaml
index 9a3cf26..59077cb 100644
--- a/fish_speech/configs/text2semantic_sft.yaml
+++ b/fish_speech/configs/text2semantic_sft.yaml
@@ -1,11 +1,11 @@
defaults:
- base
- - [email protected]: dual_ar_8_codebook_small
+ - [email protected]: dual_ar_2_codebook_small
- _self_
-project: text2semantic_sft_medium_dual_ar
+project: text2semantic_sft_dual_ar
max_length: 4096
-ckpt_path: results/text2semantic_pretrain_medium_dual_ar/checkpoints/step_000060000.ckpt
+ckpt_path: checkpoints/text2semantic-medium-v1-2k.pth
resume_weights_only: true
# Lightning Trainer
@@ -21,33 +21,28 @@ trainer:
# Dataset Configuration
tokenizer:
_target_: transformers.AutoTokenizer.from_pretrained
- pretrained_model_name_or_path: fishaudio/speech-lm-v1
+ pretrained_model_name_or_path: fishaudio/fish-speech-1
# Dataset Configuration
train_dataset:
_target_: fish_speech.datasets.text.AutoAugTextDataset
- use_data_server: false
proto_files:
- - data/protos/sft/train_Genshin.protos
- - data/protos/sft/sft.protos
+ - data/protos/sft/train
tokenizer: ${tokenizer}
max_length: ${max_length}
num_codebooks: ${model.model.config.num_codebooks}
- use_speaker: false
- phones_prob: 0.5
- interactive_prob: 0.5
+ use_speaker: 0.5
+ interactive_prob: 0.7
val_dataset:
_target_: fish_speech.datasets.text.AutoAugTextDataset
- use_data_server: false
proto_files:
- - data/protos/sft/val_Genshin.protos
+ - data/protos/sft/test
tokenizer: ${tokenizer}
max_length: ${max_length}
num_codebooks: ${model.model.config.num_codebooks}
- use_speaker: false
- phones_prob: 0.5
- interactive_prob: 0.5
+ use_speaker: 0.5
+ interactive_prob: 0.7
data:
_target_: fish_speech.datasets.text.TextDataModule