Best tts and vocoder models I found

[ShayBox]

I generated every combination of tts and vocoder model together, these are the resulting models I found with good combinations, though these still produce some bad combinations.

Here's a bash script

#!/usr/bin/env bash

declare -a text="The quick brown fox jumps over the lazy dog"

declare -a tts_models=(
    "tts_models/en/ljspeech/fast_pitch"
    "tts_models/en/ljspeech/glow-tts"
    "tts_models/en/ljspeech/speedy-speech"
    "tts_models/en/ljspeech/tacotron2-DDC_ph"
)

declare -a vocoder_models=(
    "vocoder_models/en/ljspeech/hifigan_v2"
    "vocoder_models/en/ljspeech/multiband-melgan"
    "vocoder_models/en/ljspeech/multiband-melgan"
    "vocoder_models/en/ljspeech/univnet"
    "vocoder_models/universal/libri-tts/fullband-melgan"
)

for model_name in "${tts_models[@]}"
do
    for vocoder_name in "${vocoder_models[@]}"
    do
        tts --text "${text}" --model_name "${model_name}" --vocoder_name "${vocoder_name}" --out_path "${model_name//\//-}-${vocoder_name//\//-}".wav
    done
done

Here's results, I removed any results that were just garbage sounds.
Hopefully this helps someone like me just looking for good tts without having to train models
Samples.zip

[MahdiEsrafili]

@ShayBox Thanks for your work. Can you please tell about processing time and hardware usage of these models?

[ShayBox]

to do a bulk it takes hours but for a single tts it takes longer for the module to init than to synth, its pretty quick, maybe not assistant speed but close

[MahdiEsrafili]

@ShayBox What about hardware usage? Did you consider memory usage? cpu or gpu?

[ShayBox]

I don't think it's that bad, but exact numbers I don't know, I have plenty and I'm running it on WSL Disrod so memory usage is already inflated, but if you're not training your own model it shouldn't use much to just synth the tts

[MahdiEsrafili]

Ok. If I train and use my own model, will it use more resources? why?

[ShayBox]

The best sounding in my opinion is tts_models/en/ljspeech/fast_pitch and vocoder_models/en/ljspeech/hifigan_v2
There's really no male voice models built-in

[curious-broccoli]

@gustaveracosta tts_models/en/vctk/vits has very good male and female voices (multispeaker model)

[macaujack]

Thanks @codemmmmm. Now I find p230 to be the kind of male voice I want:

tts --text "$content" --model_name "tts_models/en/vctk/vits" --out_path "$out" --speaker_idx p230

[cw72]

@macaujack How you run that code? Everytime I try to run a model without a default_vocoder like "tts_models/en/vctk/vits", I get an error: No espeak backend found. Install espeak-ng or espeak to your system.

[curious-broccoli]

try installing espeak?

[skwzrd]

Yes, apt install espeak-ng will solve it

[curious-broccoli]

Have you tried the VITS models (they also don't use vocoder models I think)?

[ShayBox]

If it's in the default list then yes, otherwise no

[knochenhans]

Hi, some of the tts_models/en/vctk/vits multispeaker indexes also sound pretty good and have a nice British accent. I’ve been using p273, p330, and p234 quite a lot.

[prakhs123]

347 335 273 <-- my favs. 273 is good

[mikebilly]

are there any good non-british english male voices?

[pncnmnp]

@mikebilly I don't think that's specified in any doc. I recommend you to fire up tts-server (something like tts-server --model_name tts_models/en/vctk/vits) and query all the speakers.

Edit: You can even write a python script to quickly generate different voices. Something like this.

[xenotropic]

They are specified in the corpus which has a speaker_ids file, but in coqui they got scrambled, see #2258. At least as of three months ago in January when I ran a script to generate all the voices, these were American accents in Coqui's VCTK-VITS: 256 M, 257 F, 270 F, 287 M, 293 F, 317 M, 360 F (may not be all)

[knoopx]

tried many of them, impressed with most them, wasn't expecting great results!

[zzzgit]

tts_models/en/ljspeech/vits--neon is the best

[devasheeshG]

please try jenny_neural

[hubertlepicki]

@devasheeshG is it the same as "tts_models/en/jenny/jenny" ? I don't see jenny_neural when I --list_models, only the other one

[devasheeshG]

Yup

[hubertlepicki]

Thanks, I tested it, it's very nice (although sounds artificial, but that's not exactly bad in my use case).

[omega3]

What is currently the best British English voice? Can anyone share their trained model? I have pretty week Intel G4560 with no GPU, so I have no chances at the moment to train.

[codepharmer]

I generated every combination of tts and vocoder model together, these are the resulting models I found with good combinations, though these still produce some bad combinations.

Here's a bash script

#!/usr/bin/env bash

declare -a text="The quick brown fox jumps over the lazy dog"

declare -a tts_models=(
    "tts_models/en/ljspeech/fast_pitch"
    "tts_models/en/ljspeech/glow-tts"
    "tts_models/en/ljspeech/speedy-speech"
    "tts_models/en/ljspeech/tacotron2-DDC_ph"
)

declare -a vocoder_models=(
    "vocoder_models/en/ljspeech/hifigan_v2"
    "vocoder_models/en/ljspeech/multiband-melgan"
    "vocoder_models/en/ljspeech/multiband-melgan"
    "vocoder_models/en/ljspeech/univnet"
    "vocoder_models/universal/libri-tts/fullband-melgan"
)

for model_name in "${tts_models[@]}"
do
    for vocoder_name in "${vocoder_models[@]}"
    do
        tts --text "${text}" --model_name "${model_name}" --vocoder_name "${vocoder_name}" --out_path "${model_name//\//-}-${vocoder_name//\//-}".wav
    done
done

Here's results, I removed any results that were just garbage sounds.
Hopefully this helps someone like me just looking for good tts without having to train models
Samples.zip

This is great! Thanks for sharing 😊