Update to the latest F5 tts version

[S-T-K]

A new update for f5 tts was released a few days ago.
It reportedly increases the quality of inference results (increase in speaker similarity and WER reduction).
Demo: https://swivid.github.io/F5-TTS_updates/
Might be worth updating.

[danielw97]

For what it's worth, just a +1 for this.
I'm not sure what needs to be changed on the alltalk side, but I've just setup on a new system to test and it downloaded the new model by default.
Currently though, with the old inference settings for v0 of f5 the audio is mostly nonsense.
Thanks for looking into this.

[danielw97]

cc @erew123 I've been doing some testing over the last few days with this, and it looks like a simple matter of updating the system/tts_engines/f5tts/available_models.json with the new filenames.
Mine looks like this, as an example.
If you would like a PR for this, let me know.

{
  "first_start_model": "f5tts_v1",
  "models": [
      {
          "model_name": "f5tts_v1",
          "folder_path": "f5tts_v1",
          "files_to_download": {
              "model_1250000.safetensors": "https://huggingface.co/SWivid/F5-TTS/resolve/main/F5TTS_v1_Base/model_1250000.safetensors",
              "vocab.txt": "https://raw.githubusercontent.com/SWivid/F5-TTS/refs/heads/main/src/f5_tts/infer/examples/vocab.txt",
              "vocos/config.yaml": "https://huggingface.co/charactr/vocos-mel-24khz/resolve/main/config.yaml",
              "vocos/pytorch_model.bin": "https://huggingface.co/charactr/vocos-mel-24khz/resolve/main/pytorch_model.bin"
          }
      },
      {
        "model_name": "e2tts_v1",
        "folder_path": "e2tts_v1",
        "files_to_download": {
            "model_1200000.safetensors": "https://huggingface.co/SWivid/E2-TTS/resolve/main/E2TTS_Base/model_1200000.safetensors",
            "vocab.txt": "https://raw.githubusercontent.com/SWivid/F5-TTS/refs/heads/main/src/f5_tts/infer/examples/vocab.txt",
            "vocos/config.yaml": "https://huggingface.co/charactr/vocos-mel-24khz/resolve/main/config.yaml",
            "vocos/pytorch_model.bin": "https://huggingface.co/charactr/vocos-mel-24khz/resolve/main/pytorch_model.bin"
        }
    }      
  ]
}

[S-T-K]

Thx so much for sharing!
Did you notice any significant differences/improvements compared to the previous version during your testing? @danielw97

[danielw97]

No problem at all, happy to share my findings.
At least to me, the new model sounds quite a bit better as compared to the original v1. Whilst it's not a massive change, it is still quite noticeable to me.
Particularly with pronunciation and accents, it seems to do better with as apposed to the previous one.

[S-T-K]

Just tried your suggested changes in the available_models.json
The new files were downloaded, but the results of inference using the new model were just noise. Works fine when I switch back to the previous f5 version. Did you make any other changes to get it running properly?
Oh and I guess you used the alltalk beta branch, right?

[S-T-K]

@danielw97
I guess that's what you meant when you said "Currently though, with the old inference settings for v0 of f5 the audio is mostly nonsense."
So maybe I just misunderstood, but were you able to generate proper audio using the new model from within Alltalk?
I looked through the changes in the f5 repo between the previous version and the new v1, and it seems like it's not just the model that changed, but many inference settings in general.

[yujiro85]

Just tried your suggested changes in the available_models.json The new files were downloaded, but the results of inference using the new model were just noise. Works fine when I switch back to the previous f5 version. Did you make any other changes to get it running properly? Oh and I guess you used the alltalk beta branch, right?

I made a backup and activated the env by start_environment.bat and installed f5 tts 1.0.3 (now they released 1.0.8) and it worked for model_1250000. But it made model_1200000 gibberish/noisy. 1250000 had some issues in 1.0.3, I started to get low quality outputs than 1200000. I don't know if they fixed it in 1.0.8.