omnivoice-tts

name	omnivoice-tts
description	Generate local text-to-speech audio with Jay's installed OmniVoice CLI. Use when Codex needs to create Chinese or multilingual speech, voiceover WAV files, short narration clips, test TTS output, or stable voice-cloned audio from a reference recording using `/Users/maxazure/jayhome/projects/OmniVoice`.

OmniVoice TTS

Use Jay's local OmniVoice install to generate WAV speech from text. The installed project is:

/Users/maxazure/jayhome/projects/OmniVoice

The preferred wrapper is:

scripts/omnivoice_tts.py

Quick Start

Generate a Chinese WAV with the default male Mandarin-style voice:

python3 /Users/maxazure/.codex/skills/omnivoice-tts/scripts/omnivoice_tts.py \
  --text "你好，Jay，这是用 OmniVoice 生成的一段中文语音。"

The script prints the output path. By default, files are saved under:

/Users/maxazure/jayhome/projects/OmniVoice/outputs/

Default Settings

Use these defaults unless the user asks otherwise:

• language: Chinese
• instruct: 男，中音调
• num_step: 16
• class_temperature: 0
• position_temperature: 0
• output format: WAV, 24 kHz mono

For a faster rough preview, pass --num-step 8.

Voice Consistency

For consistent voice across multiple generations, prefer voice cloning:

python3 /Users/maxazure/.codex/skills/omnivoice-tts/scripts/omnivoice_tts.py \
  --text "这里是要生成的新文本。" \
  --ref-audio /path/to/reference.wav \
  --ref-text "这里写参考音频中实际说的话。" \
  --output /path/to/output.wav

Keep the same ref-audio, ref-text, language, instruct, num-step, speed or duration, and temperatures for every segment. Avoid Auto Voice when the user asks for a consistent speaker.

Reference audio guidance:

• Use 3-10 seconds of clean speech.
• Use the same language as the target text when possible.
• Provide accurate ref-text; omitting it makes stability worse because ASR must infer it.
• Do not clone a real person's voice unless the user has rights or consent.

Longer Scripts

For paragraphs or article narration:

• Generate one paragraph at a time if the user needs easier review or editing.
• Use the same reference audio and settings for all segments.
• Name files in reading order, such as part_001.wav, part_002.wav.
• If exact duration matters, pass --duration and --postprocess-output false.

Validation

After generation, verify the file exists and has audio:

ffprobe -v error -show_entries stream=sample_rate,channels,duration \
  -of default=nw=1 /path/to/output.wav

For a quick waveform sanity check:

python3 - <<'PY'
import soundfile as sf, numpy as np
x, sr = sf.read("/path/to/output.wav")
print("sr", sr, "samples", len(x), "duration", len(x) / sr, "peak", float(np.max(np.abs(x))))
PY