Help setting speed using CLI with multispeech

[sv-ctrl]

Checks

• This template is only for usage issues encountered.
• I have thoroughly reviewed the project documentation but couldn't find information to solve my problem.
• I have searched for existing issues, including closed ones, and couldn't find a solution.
• I am using English to submit this issue to facilitate community communication.

Environment Details

set up using the instructions in the readme. windows 11.

Steps to Reproduce

using multi-example with CLI and .toml file

✔️ Expected Behavior

No response

❌ Actual Behavior

So first off, I love this repo and everyone involved in it. It is excellent. The documentation is excellent, and the gradio app is so easy to use.

HOWEVER. Having gone through the documentation and the apps I am having an issue that I hope can be easily solved, but the closest answer I found in the discussions is "see gradio_app advanced settings".

I am using the CLI to do multispeech, almost exactly like the example. Where I'm struggling is I need to set the speed for various speech types to be different. An example would be Country mouse speed=0.5 and Town mouse speed =0.8.

Is this possible to pass to the main infer.py with a .toml file? How would you go about this?

Additionally, if I were to fine-tune models for different speakers, could we use a different model for each speaker? Is that possible or not in the current environment?

Thanks for any help you can give!

[SWivid]

Hi @sv-ctrl , try latest pull 52c8477
You can do like:

F5-TTS/src/f5_tts/infer/examples/multi/story.toml

Line 16 in 52c8477

speed = 0.8 # will ignore global speed

Basically is to have

F5-TTS/src/f5_tts/infer/infer_cli.py

Line 336 in 52c8477

local_speed = voices[voice].get("speed", speed)

---

Personally suggest to try "Multi-Style / Multi-Speaker Generation" in infer-gradio UI, would be much easier and direct to make content.
infer-cli also fine if device not available for GUI, and you can add random_seed feature into code similarly as local_speed:

• to pass and get seed number
• then, torch.manual_seed(seed)

[sv-ctrl]

Thank you for your response!

Just to clarify your comment for myself since I am very bad at Python, I could define in the .toml file a speed for each voice, and then in the infer_cli.py file, modify it to use the local_speed as defined in the .toml file?

I'm having good lucking using the CLI instead of the GUI since my intended use case is for long (~ 1hour) dialogue with multiple voices. I'm using LLMs to sort and arrange multi-speaker dialogue and then pass it to f5tts to read. So far it's working great (other than my issues which I hope your suggestions fix!)

[SWivid]

• The local_speed feature for infer-cli is already merged into repo, you just need to pip install -U f5-tts, or redo pip install -e .
• Got your need, I think infer-gradio suits better, try multi-style feature in it~

[sv-ctrl]

Thank you for the clarification, as well as the update to the repo. I really like the gradio app, I can tell a lot of work went into making it easy to use.

Thank you for your help and prompt responses! I'll close this now.