Text-to-Speech
ONNX
GGUF
Chinese
English
onnxruntime
tts
on-device
jetson
telephony
vits
mb-istft-vits
multi-speaker
mandarin
taiwanese-mandarin
imatrix
conversational
Instructions to use Luigi/PrimeTTS with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- llama.cpp
How to use Luigi/PrimeTTS with llama.cpp:
Install (macOS, Linux)
curl -LsSf https://llama.app/install.sh | sh # Start a local OpenAI-compatible server with a web UI: llama serve -hf Luigi/PrimeTTS:F32 # Run inference directly in the terminal: llama cli -hf Luigi/PrimeTTS:F32
Install from WinGet (Windows)
winget install llama.cpp # Start a local OpenAI-compatible server with a web UI: llama serve -hf Luigi/PrimeTTS:F32 # Run inference directly in the terminal: llama cli -hf Luigi/PrimeTTS:F32
Build from source code
git clone https://github.com/ggml-org/llama.cpp.git cd llama.cpp cmake -B build cmake --build build -j --target llama-app # Start a local OpenAI-compatible server with a web UI: ./build/bin/llama serve -hf Luigi/PrimeTTS:F32 # Run inference directly in the terminal: ./build/bin/llama cli -hf Luigi/PrimeTTS:F32
Use Docker
docker model run hf.co/Luigi/PrimeTTS:F32
- Llama new
- LM Studio
- Jan
- Ollama
How to use Luigi/PrimeTTS with Ollama:
ollama run hf.co/Luigi/PrimeTTS:F32
- Unsloth Desktop
- Docker Model Runner
How to use Luigi/PrimeTTS with Docker Model Runner:
docker model run hf.co/Luigi/PrimeTTS:F32
- Lemonade
How to use Luigi/PrimeTTS with Lemonade:
Pull the model
# Download Lemonade from https://lemonade-server.ai/ lemonade pull Luigi/PrimeTTS:F32
Run and chat with the model
lemonade run user.PrimeTTS-F32
List all available models
lemonade list
- Atomic Chat
remove superseded BreezyVoice-era script
Browse files- scripts/run_v4zh_v2.sh +0 -54
scripts/run_v4zh_v2.sh
DELETED
|
@@ -1,54 +0,0 @@
|
|
| 1 |
-
#!/usr/bin/env bash
|
| 2 |
-
# v4zh-v2: retrain on DIVERSITY-EXPANDED corpus (existing 9042 + ~6500 broad-char Tatoeba/zh-TW).
|
| 3 |
-
# Tests the data-diversity hypothesis: held-out zh should drop from 0.40 toward the 0.21 seen-text floor.
|
| 4 |
-
# Keeps the proven v4 phone-level alignment. Run when teacher synth + GPU are free. GPU via $1 (default 1).
|
| 5 |
-
set -e
|
| 6 |
-
ZT=/home/luigi/jetson-tts/mossnano/zhtw8k; LOG=$ZT/v4zh_v2.log
|
| 7 |
-
TRAIN=/home/luigi/moss-train-venv/bin/python
|
| 8 |
-
GPU=${1:-1}
|
| 9 |
-
VOC=$ZT/m2_vocoder_8k/hifigan-snake_8k-latest.pt
|
| 10 |
-
exec >>"$LOG" 2>&1
|
| 11 |
-
echo "===== V4ZH_V2 START $(date) GPU=$GPU ====="
|
| 12 |
-
|
| 13 |
-
# 1) canonical-fix expansion manifest wav paths from id (idempotent, no doubling)
|
| 14 |
-
$TRAIN - <<PY
|
| 15 |
-
import json, os
|
| 16 |
-
p="$ZT/teacher_corpus_expand/manifest.jsonl"; D="$ZT/teacher_corpus_expand"
|
| 17 |
-
rows=[json.loads(l) for l in open(p) if l.strip()]
|
| 18 |
-
ok=0
|
| 19 |
-
for r in rows:
|
| 20 |
-
r["wav"]=f"{D}/{r['id']}.wav"; ok+=os.path.exists(r["wav"])
|
| 21 |
-
open(p,"w").write("\n".join(json.dumps(r,ensure_ascii=False) for r in rows)+"\n")
|
| 22 |
-
print(f"[expand manifest] {len(rows)} rows, {ok} files exist")
|
| 23 |
-
PY
|
| 24 |
-
|
| 25 |
-
# 2) v4-align the expansion clips (phone-level)
|
| 26 |
-
echo "[align expand] $(date)"
|
| 27 |
-
CUDA_VISIBLE_DEVICES=$GPU $TRAIN $ZT/align_durations_v4.py \
|
| 28 |
-
--manifest $ZT/teacher_corpus_expand/manifest.jsonl \
|
| 29 |
-
--out $ZT/m_v4_expand_align.jsonl --device cuda
|
| 30 |
-
|
| 31 |
-
# 3) abspath-fix + merge with existing v4 alignment
|
| 32 |
-
$TRAIN - <<PY
|
| 33 |
-
import json
|
| 34 |
-
R="/home/luigi/jetson-tts/"
|
| 35 |
-
def fix(p):
|
| 36 |
-
rows=[json.loads(l) for l in open(p) if l.strip()]
|
| 37 |
-
for r in rows:
|
| 38 |
-
if not r["target_audio"].startswith("/"): r["target_audio"]=R+r["target_audio"]
|
| 39 |
-
return rows
|
| 40 |
-
a=fix("$ZT/m_v4_align.jsonl"); b=fix("$ZT/m_v4_expand_align.jsonl")
|
| 41 |
-
out="$ZT/m_v4_v2_align.jsonl"
|
| 42 |
-
open(out,"w").write("\n".join(json.dumps(r,ensure_ascii=False) for r in a+b)+"\n")
|
| 43 |
-
print(f"[merged] {len(a)} + {len(b)} = {len(a)+len(b)} -> {out}")
|
| 44 |
-
PY
|
| 45 |
-
|
| 46 |
-
# 4) retrain on merged corpus (same recipe, frozen arch)
|
| 47 |
-
echo "[v4zh-v2 acoustic] $(date)"
|
| 48 |
-
cd /tmp/inflect-nano
|
| 49 |
-
CUDA_VISIBLE_DEVICES=$GPU $TRAIN -m inflect_nano.acoustic --durations-jsonl $ZT/m_v4_v2_align.jsonl \
|
| 50 |
-
--out-dir $ZT/v4zh_v2_acoustic_8k --vocoder-variant snake_8k --sample-rate 8000 \
|
| 51 |
-
--steps 60000 --batch-size 16 --lr 2e-4 --max-frames 1000 --en-upsample 2 \
|
| 52 |
-
--vocoder-checkpoint $VOC --vocoder-mel-weight 1.0 \
|
| 53 |
-
--save-interval 5000 --log-interval 200 --device cuda
|
| 54 |
-
echo "===== V4ZH_V2 TRAIN DONE $(date) ====="
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|