mimo-tts-wav

name	mimo-tts-wav
description	Generate a spoken or sung WAV audio file using MiMo TTS. USE THIS SKILL whenever the user asks the model to speak, say something aloud, read text out, sing a song, or produce any audio/voice output — even if no explicit destination is mentioned. Save the result to the current working directory as a .wav file (default name: output.wav) unless the user specifies a different path. Supports any expressive style and custom voice cloning.
applyTo	**

MiMo TTS → WAV File

Convert text to speech using the MiMo TTS API and save the result as a standard WAV file (24 kHz, 16-bit PCM, mono) to any local path you choose.

---

When to Use

Situation	Use this skill?
User says "说..." / "读出来" / "大声念" / "speak" / "say"	✅ Yes
User says "唱..." / "唱个歌" / "唱歌" / "sing"	✅ Yes
User asks for any voice/audio output	✅ Yes
No output path given — save to $PWD/output.wav	✅ Yes
Need to send voice directly to Feishu	❌ Use mimo-tts-voice instead
Need to post-process audio (ffmpeg, SoX …)	✅ Yes — pipe the output

---

Prerequisites

Requirement	Details
jq	JSON builder. Install: brew install jq
python3	Standard library only (json, base64, struct). Pre-installed on macOS.
curl	Pre-installed on macOS.

---

Usage

bash scripts/tts_to_wav.sh "<text>" "<output.wav>" [style] [voice_sample.wav]

Arguments

#	Name	Required	Description
1	text	✅	Text to synthesize. Apply TTS normalization rules before passing. Singing mode: pass a song name (looked up in sing0301_dict.json) or LYRICS:<your lyrics> to supply lyrics directly.
2	output_path	✅	Destination path for the WAV file. Parent directory must exist.
3	style	optional	Any natural-language style phrase (see Style Guide below). Pass 唱歌 to enable singing mode.
4	voice_sample	optional	Path to a WAV reference clip (24 kHz, 16-bit, mono, 5–15 s) for voice cloning. Omit to use the preset voice mimo_default.

Environment Variables

Variable	Default	Description
MIMO_VOICE_SAMPLE	—	Fallback voice sample path (used when voice_sample arg is omitted)
MIMO_API_BASE_URL	https://api.xiaomimimo.com/v1	Override API base URL (/chat/completions is appended automatically)
MIMO_TTS_MODEL	mimo-v2.5-tts	Override model name
MIMO_SING_DICT	../../sing0301_dict.json (relative to script)	Path to sing0301_dict.json song lyrics dictionary

---

Step-by-Step

1. Normalize the text — Apply TTS normalization (see section below) so numbers, symbols and formatting render as natural speech.

2. Choose a voice

   • Preset: omit voice_sample → uses mimo_default
   • Custom clone: pass a 5–15 s WAV clip as the 4th argument, or set MIMO_VOICE_SAMPLE

3. Choose a style (optional) — pass any descriptive phrase as the 3rd argument. Omit for neutral speech.

4. Run the script:

   bash scripts/tts_to_wav.sh "要说的内容" "/path/to/out.wav"
   

5. Check the output — the script exits 0 and prints OK: <N> bytes written to <path> on success. Any error is printed to stderr with a non-zero exit code.

---

Examples

# Minimal — preset voice, no style
bash scripts/tts_to_wav.sh "今天天气不错" /tmp/hello.wav

# With a mood style
bash scripts/tts_to_wav.sh "恭喜你获得了一等奖！" /tmp/congrats.wav "开心激动"

# Slow delivery for an accessibility recording
bash scripts/tts_to_wav.sh "请系好安全带。" /tmp/safety.wav "语速慢 清晰"

# Custom voice clone + style
bash scripts/tts_to_wav.sh "大家好，我是小明。" /tmp/xiaoming.wav "热情" ~/voices/xiaoming.wav

# Singing — look up lyrics by song name (requires sing0301_dict.json)
bash scripts/tts_to_wav.sh "两只老虎" /tmp/song.wav "唱歌"

# Singing — supply lyrics directly with LYRICS: prefix
bash scripts/tts_to_wav.sh "LYRICS:两只老虎，两只老虎，跑得快，跑得快……" /tmp/song.wav "唱歌"

# Pipe straight into ffmpeg to get MP3
bash scripts/tts_to_wav.sh "你好" /tmp/tmp.wav && \
  ffmpeg -i /tmp/tmp.wav /tmp/out.mp3

---

Style Guide

style accepts any natural-language phrase — the model interprets it freely. There are no fixed values.

Category	Examples
Emotion	开心 / 悲伤 / 生气 / 平静
Delivery	语速慢 / 语速快 / 悄悄话 / 清晰有力
Character	像个大将军 / 像个小孩 / 孙悟空 / 林黛玉
Dialect	东北话 / 四川话 / 台湾腔 / 粤语
Combinations	慵懒 刚睡醒 / 撒娇 夹子音 / 深情款款 语速慢
Singing	唱歌 — text is a song name (looked up in sing0301_dict.json) or LYRICS:<lyrics>

Omit the style argument entirely for neutral, natural-sounding speech.

---

TTS Text Normalization

Preprocess the input text before passing it to the script. Replace unspoken symbols with natural spoken Chinese (or English where appropriate).

Numbers

Input	Spoken
3	三
3.14	三点一四
1/3	三分之一
14:30	下午两点半
95%	百分之九十五
2024 (year)	二零二四年

Common Symbols

Symbol	Spoken
+	加
-	减
× / *	乘以
÷ / /	除以
=	等于
> / <	大于 / 小于
%	百分之…
~	大约
... / …	等等
#	井号 (or 第, by context)

Formatting Artifacts

• Remove markdown (**bold**, # heading, `code`, - bullet)
• Convert numbered lists to prose: "有三点，第一……第二……第三……"
• Convert tables to sentences describing key values

---

Output Format

The script always writes a standard RIFF WAV file:

Property	Value
Format	WAV (RIFF)
Sample rate	24 000 Hz
Bit depth	16-bit PCM
Channels	Mono

If the API returns raw PCM instead of WAV, the script automatically adds the correct RIFF header before writing to disk.

---

Error Reference

Exit code	Meaning	Fix
1	Missing required argument	Provide text and output_path
1	Voice sample file not found	Check the path passed as arg 4
2	API HTTP error	Check network access; API key is resolved automatically
2	Unexpected response shape	Inspect stderr for raw API response
3	Output file empty / not written	Check that the parent directory exists and is writable