「TTS.cpp」を試す

git clone https://github.com/mmwillet/TTS.cpp && cd TTS.cpp

git clone -b support-for-tts https://github.com/mmwillet/ggml.git

export ESPEAK_INSTALL_DIR=/opt/homebrew/opt/espeak-ng
cmake -B build
cmake --build build --config Release

mkdir models
pushd models

# Parler-TTS
# miniとlargeがあり、それぞれF32、Q5がある。
# ここではlarge・Q5をダウンロード
wget --content-disposition https://huggingface.co/mmwillet2/Parler_TTS_GGUF/resolve/main/Parler_TTS_large_Q5.gguf?download=true

# Dia
# Q4、Q5，Q8、F16、F32がある。
# またそれに`_DAC_F16`がついているものは16ビットDACエンコーダでエンコーダされたバージョン。
# 今回はQ5・16ビットDACエンコーダ版を使う
wget --content-disposition  https://huggingface.co/mmwillet2/Dia_GGUF/resolve/main/Dia_Q5_DAC_F16.gguf?download=true

# Kokoro-TTS
# Espeak-ng対応版・非対応版がある様子。
# 英語だけならEspeak-ng非対応のもので良さそうだが、日本語でも使いたいのと一部記号を上手く処理できないみたいなので、Espeak-ng対応版を。
wget --content-disposition https://huggingface.co/mmwillet2/Kokoro_GGUF/resolve/main/Kokoro_espeak.gguf?download=true

popd

./build/bin/cli --help

--temperature (-t):
    The temperature to use when generating outputs. Defaults to 1.0.
--repetition-penalty (-r):
    The by channel repetition penalty to be applied the sampled output of the model. defaults to 1.0.
--top-p (-tp):
    (OPTIONAL) the sum of probabilities to sample over. Must be a value between 0.0 and 1.0. Defaults to 1.0.
--n-threads (-nt):
    The number of cpu threads to run generation with. Defaults to hardware concurrency. If hardware concurrency cannot be determined then it defaults to 1.
--topk (-tk):
    (OPTIONAL) When set to an integer value greater than 0 generation uses nucleus sampling over topk nucleaus size. Defaults to 50.
--max-tokens (-mt):
    (OPTIONAL) The max audio tokens or token batches to generate where each represents approximates 11 ms of audio. Only applied to Dia generation. If set to zero as is its default then the default max generation size. Warning values under 15 are not supported.
--use-metal (-m):
    (OPTIONAL) Whether to use metal acceleration
--no-cross-attn (-ca):
    (OPTIONAL) Whether to not include cross attention
--vad (-va):
    (OPTIONAL) whether to apply voice inactivity detection (VAD) and strip silence form the end of the output (particularly useful for Parler TTS). By default, no VAD is applied.
--play:
    (OPTIONAL) Whether to play back the audio immediately instead of saving it to file.
--model-path (-mp):
    (REQUIRED) The local path of the gguf model file for Parler TTS mini or large v1, Dia, or Kokoro.
--prompt (-p):
    (REQUIRED) The text prompt for which to generate audio in quotation markers.
--save-path (-sp):
    (OPTIONAL) The path to save the audio output to in a .wav format. Defaults to TTS.cpp.wav
--conditional-prompt (-cp):
    (OPTIONAL) A distinct conditional prompt to use for generating. If none is provided the preencoded prompt is used. '--text-encoder-path' must be set to use conditional generation.
--text-encoder-path (-tep):
    (OPTIONAL) The local path of the text encoder gguf model for conditional generaiton.
--voice (-v):
    (OPTIONAL) The voice to use to generate the audio. This is only used for models with voice packs.
--espeak-voice-id (-eid):
    (OPTIONAL) The espeak voice id to use for phonemization. This should only be specified when the correct espeak voice cannot be inferred from the kokoro voice ( see MultiLanguage Configuration in the README for more info).

./build/bin/cli \
    --model-path models/Parler_TTS_large_Q5.gguf \
    --prompt "Good morning. It's a beautiful day today. On days like this, I feel like going to the horse races." \
    --save-path tts-cpp-parlertts-sample.wav

./build/bin/cli \
    --model-path models/Dia_Q5_DAC_F16.gguf \
    --prompt "[S1] Good morning. It's a beautiful day today. On days like this, I feel like going to the horse races." \
    --save-path tts-cpp-dia-sample.wav

./build/bin/cli \
    --model-path models/Kokoro_espeak.gguf \
    --prompt "Good morning. It's a beautiful day today. On days like this, I feel like going to the horse races." \
    --save-path tts-cpp-kokoro-tts-sample.wav

./build/bin/cli \
    --model-path models/Kokoro_espeak.gguf \
    --prompt "おはようございます。今日はいいお天気ですね。こんな日は競馬に行きたくなりますね。" \
    --save-path tts-cpp-kokoro-tts-sample-ja.wav \
    --voice "jf_alpha"

espeak-ng -v jpx/ja -x "おはようございます。今日はいいお天気ですね。こんな日は競馬に行きた くなりますね。"

,ohaj,oug,ozaim'asu
k_j,ouh,ai:,oteNk,ides'une
k,onnaC,ihak,eib,ani:k,itak,unar`,imas'une

from misaki import ja, en

g2p = ja.JAG2P()

phonemes = g2p("おはようございます。きょうはいいおてんきですね。こんなひはけいばにいきたくなりますね。")
print(phonemes)

('ohajoː ɡoʣaimasɨ kʲoː βa iː o teŋkʲi desɨ ne konna çiβa keːba ɲi ikʲi takɯ naɾʲimasɨ ne', None)

./build/bin/cli \
    --model-path models/Dia_Q5_DAC_F16.gguf \
    --prompt "[S1] Good morning. It's a beautiful day today. [S2] Yep. On days like this, I feel like going to the horse races. [S1] (laughs) Definitely." \
    --save-path tts-cpp-dia-sample2.wav \
    --topk 35 \
    --temperature 1.3

uv venv -p 3.12.9
uv pip install pip
source .venv/bin/activate

cd py-gguf
pip install -r requirements.txt

python ./convert_t5_encoder_to_gguf --save-path ../models/t5-encoder-large.gguf --large-model

WARNING:parler_tts.modeling_parler_tts:Flash attention 2 is not installed
INFO:gguf.gguf_writer:gguf: This GGUF file is for Little Endian only
INFO:T5Encoder:Preparing tensors for GGUF file, t5-encoder-large.gguf.
config.json: 100%|█████████████████████████████████████████████████████████████████| 7.72k/7.72k [00:00<00:00, 7.89MB/s]
model.safetensors.index.json: 100%|█████████████████████████████████████████████████| 94.5k/94.5k [00:00<00:00, 618kB/s]
Downloading shards:   0%|                                                                         | 0/2 [00:00<?, ?it/s]
model-00001-of-00002.safetensors:  16%|███████                                      | 776M/4.98G [01:23<07:32, 9.31MB/s]
(snip)
INFO:T5Encoder:Preparing configuration for GGUF file, ../models/t5-encoder-large.gguf.
INFO:T5Encoder:Beginning to write to GGUF file, ../models/t5-encoder-large.gguf.
INFO:gguf.gguf_writer:Writing the following files:
INFO:gguf.gguf_writer:../models/t5-encoder-large.gguf: n_tensors = 221, total_size = 4.9G
Writing: 100%|██████████████████████████████████████████████████████████████████| 4.91G/4.91G [00:04<00:00, 1.02Gbyte/s]
INFO:T5Encoder:Finishing GGUF converstion process for file, ../models/t5-encoder-large.gguf.

cd ..

./build/bin/cli \
    --model-path models/Parler_TTS_large_Q5.gguf \
    --prompt "I am saying some words" \
    --text-encoder-path models/t5-encoder-large.gguf \
    --conditional-prompt "male deep voice" \
    --save-path tts-cpp-parlertts-sample2.wav

Segmentation fault: 11