12s limit warning

[RemainIndoors1]

I don't know if this is a bug per se, but in case anybody else runs into this issue, it's probably good to be aware of at least. If you try to use reference audio that is longer than 12s, there is logic in f5_tts/infer/utils_infer.py that trims the audio down to 12 seconds. The problem comes into play when you also specify a transcription for your audio, because it won't trim the transcription to match what your new audio clip says after the trim, which leads to some really odd behavior with the generated audio.

One solution could be, if it Has to trim the audio to 12s for some reason, it could force calling the transcribe method for your new audio length, but to be on the safe side, you should make sure your sample audio is less than 12 seconds.

[polarathene]

For added context, this is documented here: https://github.com/yl4579/DMOSpeech2/blob/main/src/f5_tts/infer/README.md

Currently support 30s for a single generation, which is the total length (same logic if fix_duration) including both prompt and output audio. However, infer_cli and infer_gradio will automatically do chunk generation for longer text. Long reference audio will be clip short to ~12s.

To avoid possible inference failures, make sure you have seen through the following instructions.

• Use reference audio <12s and leave proper silence space (e.g. 1s) at the end. Otherwise there is a risk of truncating in the middle of word, leading to suboptimal generation.
• Uppercased letters (best with form like K.F.C.) will be uttered letter by letter, and lowercased letters used for common words.
• Add some spaces (blank: " ") or punctuations (e.g. "," ".") to explicitly introduce some pauses.
• If English punctuation marks the end of a sentence, make sure there is a space " " after it. Otherwise not regarded as when chunk.
• Preprocess numbers to Chinese letters if you want to have them read in Chinese, otherwise in English.
• If the generation output is blank (pure silence), check for FFmpeg installation.
• Try turn off use_ema if using an early-stage finetuned checkpoint (which goes just few updates).

---

Additionally it's unlikely this project author will attend to that since for the most part, the src/f5-tts source is synced with upstream at the 1.1.7 release (July). However there doesn't appear to be any change to the same document upstream, but I would suggest raising an issue there if you'd like to get the bug addressed.