Ecosyste.ms: Timeline
Browse the timeline of events for every public repo on GitHub. Data updated hourly from GH Archive.
SWivid created a comment on an issue on SWivid/F5-TTS
So you are training from scratch or finetuning on pretrained model. A probably reason is: the low loss is for overfitting if you use small batchsize.
SWivid created a comment on an issue on SWivid/F5-TTS
@kmn1024 Ah yes, that makes sense. Maybe the best way is to filter out null audio samples or we just remove the contraint <0.3?
SWivid created a comment on an issue on SWivid/F5-TTS
@trunks970 some similar issues, try online space demo in huggingface/modelscope. Probably with whisper asr failure or network stuffs...
lpscr created a comment on a pull request on SWivid/F5-TTS
@SWivid also about the vocab new from dataset i think be good to extend with vocab for Emilia_ZH_EN_pinyin for example if there is new symbols not in let to add this from train from the scrat...
jpgallegoar created a comment on a pull request on SWivid/F5-TTS
@SWivid This speeds inference a lot, good idea
jpgallegoar pushed 3 commits to main SWivid/F5-TTS
jpgallegoar closed a pull request on SWivid/F5-TTS
Added audio caching for faster inference
Added audio caching in utils_infer so if the same audio is used twice, it will not be transcribed again.
lpscr opened a pull request on SWivid/F5-TTS
new update gradiof finetune
just another great update fix some stuf first add create vocan from the dataset and you can see  s...
jpgallegoar closed a pull request on SWivid/F5-TTS
Added audio cashing for faster inference
Added audio caching in utils_infer so if the same audio is used twice, it will not be transcribed again.
jpgallegoar opened a pull request on SWivid/F5-TTS
Added audio cashing for faster inference
Added audio caching in utils_infer so if the same audio is used twice, it will not be transcribed again.
huutuongtu opened an issue on SWivid/F5-TTS
Model generates audio with random content during inference
Hi, thanks for the great work. I trained the model on my own dataset (Vietnamese, 1 speaker, 1.5 hours, just for testing) for about 70k steps. In [this comment,](https://github.com/SWivid/F5-TTS/is...
kmn1024 created a comment on an issue on SWivid/F5-TTS
Could the following happen? Suppose we have these examples with these frames lengths and durations. ex1: 100, 2.0 ex2: 50, 1.0 ex3: 10, 0.2 ex4: 100, 2.0 Suppose frames_threshold is 101. The...
trunks970 created a comment on an issue on SWivid/F5-TTS
> Use lower case as we suggested in readme, or you are telling model to read letter by letter Also check if reference audio uploaded correctly, will show waveform if so done both unfortunately n...
kmn1024 created a comment on an issue on SWivid/F5-TTS
What I mean to suggest is, suppose we have these examples with these durations ex1, 100 ex2, 1 ex2, 1 ex2, 1
SWivid created a comment on a pull request on SWivid/F5-TTS
@lpscr the `dev` branch is from this repo. updates is mainly for fair comparison with other models with different train set. as the trainset is of different size, so academic comparison general...
jpgallegoar created a comment on a pull request on SWivid/F5-TTS
Thank you for the inputs, I also thought about the extra loading time. I will look into it and update.
SWivid created a comment on a pull request on SWivid/F5-TTS
@jpgallegoar Hi, very cool chat! I just test it, may need few tweaks with your help: 1. will the ref_text saved and not calling asr pipeline to do transcription again if not ref_audio not chang...