ttscn — Multi-Platform Chinese TTS Skill

Overview

Generate natural speech audio from text. 14 backends — 11 China-friendly clouds plus 3 international (ElevenLabs / OpenAI / Google).

#	Backend	Cost	Key strength
1	Edge TTS (default)	Free	No API key, works everywhere
2	Doubao (ByteDance)	~1 RMB/10K	Best Chinese naturalness (9/10)
3	CosyVoice (Alibaba)	~0.2 RMB/1K	Fast streaming, flexible
4	Qwen3-TTS (Alibaba)	~1 RMB/10K	10 languages, instruction control
5	StepFun (阶跃星辰)	Low cost	OpenAI-compatible, ~10s cloning
6	GLM-TTS (Zhipu)	Low cost	Emotion control, simple REST
7	Azure (Microsoft)	~1 USD/M chars	Enterprise SSML, eastasia
8	Tencent Cloud	0.75 RMB/10K	Lowest cost, 380+ voices
9	Baidu AI	Flexible	30+ voices, emotion + dialects
10	MiniMax	~$0.10/1K	Best quality, 300+ voices, cloning
11	iFlytek Xunfei	~2 RMB/10K	MOS 4.8, 500+ voices, pro grade
12	ElevenLabs	Paid tiers (from $5/mo)	Top voice quality, instant cloning
13	OpenAI TTS	~$15-30/M chars	6 voices, multilingual, simple REST
14	Google Cloud TTS	~$16/M chars (free tier)	220+ voices, 40+ languages

New in 1.4–1.6: word-level timestamps (edge/azure/doubao/minimax/cosyvoice — best-effort, degrades to no boundaries), [PAUSE:x] + sound-tag markers (all platforms), --phonemes pronunciation overrides (azure/minimax). New in 1.7: --json flag (JSON envelope independent of --format), idempotency hits report cached: true and re-synthesize if the cached audio was deleted, --input f out.wav positional fixed, chunker never splits inside [PAUSE:x]. New in 1.8: MiniMax defaults to speech-2.8-hd (sound tags voiced out of the box), CosyVoice v3.5-flash supported (custom/cloned voices only — presets need cosyvoice-v3-flash). New in 1.9: Qwen3-TTS (DashScope, reuses DASHSCOPE_API_KEY), StepFun (reuses STEP_API_KEY) and GLM-TTS (reuses ZHIPUAI_API_KEY) backends — 14 backends total, all three reuse keys you already have.

Cross-platform: Windows, macOS, Linux

All paths in this document are relative to this skill's root directory (the directory containing this SKILL.md) — resolve them against it.

When to Use This Skill

Automatically activate this skill when:

• User wants to convert Chinese text to speech audio
• Generating voice narration or voiceover for videos
• Creating audiobook or podcast audio from text
• User asks to compare TTS providers, choose a TTS backend, or see what voices are available
• User asks about TTS pricing, features, or which provider supports cloning/SSML/dialects
• User mentions any of: TTS, text-to-speech, 语音合成, 文字转语音, Edge TTS, Doubao TTS, CosyVoice, 火山引擎, 阿里云语音, Azure TTS, 腾讯云TTS, 百度语音, MiniMax, 讯飞语音, ElevenLabs, OpenAI TTS, Google Cloud TTS
• User needs word-level timestamps/subtitles, pause control, or fixing mispronounced Chinese characters (多音字)
• Any task where Chinese text-to-speech would be helpful

Provider Comparison Page

When the user wants to browse, compare, or choose a TTS provider, ALWAYS open the local HTML comparison page in their browser FIRST — it's a visual, filterable table that is much faster to scan than reading text output.

# Cross-platform (path relative to this skill's directory)
python3 -m webbrowser docs/providers.html

# …or the platform-native opener: open (macOS) / xdg-open (Linux) / start (Windows)

The comparison page includes:

• Filterable table — filter by free, SSML, voice cloning, streaming, dialects, multilingual
• Per-provider detail panels — cost, max chars/duration, clone method, emotion, languages
• Voice cards — recommended voices with style descriptions and best-use labels
• API key links — direct links to each provider's console for key acquisition

This page is auto-generated from data/providers.json (the single source of truth for all provider/voice data). Run python3 scripts/build_docs.py to regenerate it after editing the JSON. The same data is queryable on the CLI via python3 scripts/tts.py schema backends|voices (see Schema Introspection).

After opening the page, ask the user which backend and voice they'd like to use, then proceed to Step 2.

Workflow

Step 0 — Show the comparison page (when comparing/choosing)

If the user is browsing, comparing providers, or unsure which backend to use, open docs/providers.html as shown above. Let them explore, then ask which backend + voice they want.

Step 1 — Understand the request

Clarify what the user needs:

• Text: inline text or a file? Short or long-form?
• Voice style: male/female, young/mature, warm/energetic? (see voice guide below)
• Speed: normal, faster (+10-20%), slower (-10-20%)?
• Format: WAV (lossless) or MP3 (compressed)?

Step 2 — Pick a backend & voice

Choose based on the use case (see Backend Selection Guide). Default to Edge TTS with zh-CN-XiaoxiaoNeural (female, warm, standard) if unsure. Mention your choice.

Step 3 — Synthesize

Run scripts/tts.py with the text and chosen options.

Step 4 — Report

Confirm: output path, file size, audio duration.

Backend Selection Guide

Quick Pick

Use case	Backend	Voice	Why
Default / general	edge	zh-CN-XiaoxiaoNeural	Free, no setup
Short video / Douyin	doubao	BV001_streaming	Native short-video style
Audiobook / long-form	cosyvoice	longxiaochun_v3	Fast synthesis, natural
Enterprise / SSML	azure	zh-CN-XiaoxiaoNeural	Rich prosody control
Bulk / lowest cost	tencent	101001	0.75 RMB/10K chars
Emotion / dialects	baidu	3 or 4	Emotion synthesis, Cantonese
Best quality / cloning	minimax	female-shaonv	speech-2.8-hd, voice design
Education / pro	xunfei	xiaoyan	MOS 4.8, 500+ voices
Male narration	edge	zh-CN-YunxiNeural	Energetic male voice
Documentary	azure	zh-CN-YunyangNeural	Deep, professional male
Children's content	edge	zh-CN-XiaomengNeural	Bright, youthful female
Cost-sensitive	edge	zh-CN-XiaoxiaoNeural	Completely free
English, top quality	elevenlabs	21m00Tcm4TlvDq8ikWAM (Rachel)	Best-in-class English voices
English, simple/cheap	openai	alloy	tts-1-hd, one env var
English, enterprise	google	en-US-Neural2-F	220+ voices, free tier
Multilingual / instruct	qwen	Cherry	10 languages, natural-language style control, reuses your DASHSCOPE key
Marketing / fast	stepfun	cixingnansheng	OpenAI-compatible, reuses your STEP key
Simple / emotion	zhipu	tongtong	Minimal REST, reuses your ZHIPUAI key

Full capability & voice data

The complete capability matrix (cost, max chars/duration per chunk, SSML, cloning method + cost, emotion, dialects, languages, streaming, setup difficulty) and the per-backend voice lists live in one place: data/providers.json. View them via:

• The comparison page (docs/providers.html) — best for humans
• python3 scripts/tts.py schema backends --full — full capability fields per backend
• python3 scripts/tts.py schema voices — every voice preset with style descriptions
• python3 scripts/tts.py --list — human-readable terminal summary

Do not maintain copies of these tables elsewhere — regenerate from the JSON.

Voice Cloning (clone command)

Create a custom voice from reference audio, store it under a name, then use the name anywhere --voice is accepted. Built-in for minimax (local file OK, 10s-5min audio, paid: ~$1.5/voice global site or ¥9.9 on first use China site; a new clone is TEMPORARY until its first real synthesis — use it within 7 days [global site] / 48 h [China site] of creation or MiniMax deletes it, previews don't count; permanent after first use) and cosyvoice (enrollment free, audio must be a PUBLIC http(s) URL, 10-20s, voice expires after 1 year unused).

# MiniMax — local file, paid, must confirm with --yes
python3 scripts/tts.py clone create --platform minimax --audio my_voice.wav --name myvoice --yes

# CosyVoice — free, but --audio must be a public URL; --target-model must
# match the synthesis model (default: $COSYVOICE_MODEL or cosyvoice-v3-flash).
# v3.5-flash has NO preset voices — it needs a custom/cloned voice created with
# the same model (voice IDs are not interchangeable across models).
python3 scripts/tts.py clone create --platform cosyvoice --audio https://example.com/my.wav --name myvoice

# Manage
python3 scripts/tts.py clone list
python3 scripts/tts.py clone delete --name myvoice [--remote]   # --remote: cosyvoice only

# Use it — the stored name resolves to the platform voice_id automatically
python3 scripts/tts.py "用我的声音说这句话" out.wav --platform minimax --voice myvoice

Rules the agent MUST follow:

• MiniMax creation is paid — never run clone create --platform minimax without the user's explicit confirmation (the CLI enforces --yes).
• Only clone the user's own voice or one they are authorized to use — both platforms contractually prohibit cloning third parties without consent.
• Reference audio: clean single-speaker speech, no BGM, 10-20s is ideal.
• Named voices live in ~/.ttscn.json under cloned_voices.

Other platforms (Doubao/Tencent/Baidu/Xunfei/Azure) support cloning via their consoles — the resulting voice id also works as a plain --voice.

Usage

Basic Usage

# Default (Edge TTS, free, Xiaoxiao voice)
python3 scripts/tts.py "你好世界" output.wav

# Specific voice
python3 scripts/tts.py --voice zh-CN-YunxiNeural "欢迎收听今天的节目" welcome.wav

# Specific backend
python3 scripts/tts.py --platform doubao "今天天气真好" weather.wav
python3 scripts/tts.py --platform minimax "高品质语音合成" hq.wav

# Adjust speed
python3 scripts/tts.py --rate +15% "快速播报" fast.wav
python3 scripts/tts.py --rate -10% "慢速朗读" slow.wav

From File

python3 scripts/tts.py --input script.txt output.wav

Output Format

# MP3 output (compressed, smaller file)
python3 scripts/tts.py --format mp3 "你好" hello.mp3

# JSON envelope + MP3 audio at the same time
python3 scripts/tts.py --json --format mp3 "你好" hello.mp3

Preview (Dry Run)

# Preview without making API call — no package installs needed
python3 scripts/tts.py --dry-run "这是一段测试文本"

List Options

python3 scripts/tts.py --list

Expressiveness Markers

Input text may contain markers on any platform — they are rendered natively where supported and stripped everywhere else (never read aloud). The chunker never splits inside a [...] marker.

Marker	Syntax	azure	minimax	all other platforms
Pause	[PAUSE:x] — x = seconds, 0.01-99.99	<break time="xs"/> (SSML)	<#x#>	stripped
Sound tags	(laughs) (chuckle) (sighs) (breath) (inhale) (exhale) (coughs)	stripped	voiced only if MINIMAX_MODEL starts with speech-2.8 (else stripped + stderr warning)	stripped

python3 scripts/tts.py --platform azure \
  "大家好。[PAUSE:0.8] 今天我们聊一个新话题。" out.wav

MINIMAX_MODEL=speech-2.8-hd python3 scripts/tts.py --platform minimax \
  "这也太好笑了 (laughs) 好，我们继续。" out.wav

Pronunciation Overrides (--phonemes)

Fix polyphonic Chinese characters (多音字) with a JSON dict mapping words to space-separated pinyin — tone-numbered (hang2 zhang3) or tone-marked (háng zhǎng). Keys starting with _ are comments.

{
  "_comment": "pronunciation overrides for bank-themed script",
  "行长": "hang2 zhang3",
  "重庆": "chóng qìng"
}

python3 scripts/tts.py --platform azure --phonemes phonemes.json \
  "行长在重庆开会。" out.wav

Per-platform: azure → SSML <phoneme alphabet="sapi"> tags; minimax → inline pinyin annotations like 重(chong2)庆(qing4) (applied before chunking so the annotation counts toward the chunk budget); all other platforms silently ignore the flag.

Requirements

# Core (always needed)
pip install edge-tts  # For Edge (default, free)

# Optional backends — install only what you use
pip install dashscope                              # CosyVoice
pip install requests                               # Doubao, MiniMax, ElevenLabs, OpenAI, Google
pip install azure-cognitiveservices-speech          # Azure
pip install tencentcloud-sdk-python-tts             # Tencent Cloud
pip install baidu-aip chardet                       # Baidu AI
pip install websocket-client                        # Xunfei

System requirement: ffmpeg

Environment Variables

# Global defaults (optional)
export TTS_BACKEND="edge"
export TTS_VOICE="zh-CN-XiaoxiaoNeural"
export TTS_RATE="+5%"
export TTS_FORMAT="wav"              # wav | mp3 | json (json = JSON envelope mode)

# Backend tuning (optional)
export MINIMAX_MODEL="speech-2.8-hd"       # default; sound tags need speech-2.8-*
export MINIMAX_GROUP_ID=""                 # required by some MiniMax accounts
export COSYVOICE_MODEL="cosyvoice-v3-flash"  # or cosyvoice-v3.5-flash (custom/cloned voices only)

# ByteDance Volcano Ark (Doubao)
# v3 (recommended, no appid): API key from the new console (Ark API Key page)
export VOLCENGINE_API_KEY="your_api_key"
export VOLCENGINE_RESOURCE_ID="seed-tts-2.0"   # optional; seed-tts-1.0 / seed-icl-2.0 also valid
# v1 (legacy, appid + token): only when VOLCENGINE_API_KEY is unset
export VOLCENGINE_APPID="your_app_id"
export VOLCENGINE_ACCESS_TOKEN="your_token"

# Alibaba DashScope (CosyVoice + Qwen3-TTS)
export DASHSCOPE_API_KEY="your_api_key"
export QWEN_TTS_MODEL="qwen3-tts-flash"   # or qwen3-tts-instruct-flash (instructions control)
export QWEN_TTS_LANGUAGE=""              # optional: Chinese / English / ... ; unset = Auto
export QWEN_TTS_INSTRUCTIONS=""          # optional: natural-language style, instruct models only

# StepFun (阶跃星辰)
export STEP_API_KEY="your_api_key"
export STEPFUN_TTS_MODEL="step-tts-mini"   # or step-tts-2 / stepaudio-2.5-tts

# Zhipu (智谱)
export ZHIPUAI_API_KEY="your_api_key"
export ZHIPU_TTS_MODEL="glm-tts"

# Microsoft Azure
export AZURE_SPEECH_KEY="your_key"
export AZURE_SPEECH_REGION="eastasia"
export TTS_STYLE="gentle"              # optional: mstts:express-as style; unset = plain prosody

# Tencent Cloud
export TENCENT_SECRET_ID="your_secret_id"
export TENCENT_SECRET_KEY="your_secret_key"

# Baidu AI
export BAIDU_APP_ID="your_app_id"
export BAIDU_API_KEY="your_api_key"
export BAIDU_SECRET_KEY="your_secret_key"

# MiniMax
export MINIMAX_API_KEY="your_api_key"

# iFlytek Xunfei
export XUNFEI_APP_ID="your_app_id"
export XUNFEI_API_KEY="your_api_key"
export XUNFEI_API_SECRET="your_api_secret"

# ElevenLabs (international)
export ELEVENLABS_API_KEY="your_api_key"
export ELEVENLABS_MODEL="eleven_multilingual_v2"   # optional, this is the default

# OpenAI TTS (international)
export OPENAI_API_KEY="your_api_key"
export OPENAI_TTS_MODEL="tts-1-hd"                 # optional, this is the default

# Google Cloud TTS (international)
export GOOGLE_TTS_API_KEY="your_api_key"
export GOOGLE_TTS_LANGUAGE="en-US"                 # optional, auto-derived from voice name

Get API Keys:

• Volcano Ark: https://console.volcengine.com/ark/region:ark+cn-beijing/apikey
• DashScope: https://bailian.console.aliyun.com/
• Azure: https://portal.azure.com/
• Tencent Cloud: https://console.cloud.tencent.com/tts
• Baidu AI: https://console.bce.baidu.com/ai/#/ai/speech/overview
• MiniMax: https://platform.minimaxi.com
• Xunfei: https://www.xfyun.cn
• ElevenLabs: https://elevenlabs.io/app/settings/api-keys
• OpenAI: https://platform.openai.com/api-keys
• Google Cloud: https://console.cloud.google.com/apis/credentials

Config File (Optional)

Create ~/.ttscn.json for personal defaults, or .ttscn.json in a project directory:

{
  "backend": "minimax",
  "voice": "female-shaonv",
  "rate": "+10%"
}

Priority (highest first):

1. CLI arguments (--platform, --voice, --rate)
2. Environment variables (TTS_BACKEND, TTS_VOICE, TTS_RATE)
3. Project config (.ttscn.json in current directory)
4. User config (~/.ttscn.json)
5. Built-in defaults

Examples

Quick Narration (Free, Zero Setup)

python3 scripts/tts.py \
  "人工智能正在改变我们的生活方式，从智能助手到自动驾驶，技术革新无处不在。" \
  ai_narration.wav

Douyin Style Short Video Voice

python3 scripts/tts.py \
  --platform doubao --voice BV001_streaming --rate +10% \
  "家人们，今天给大家推荐一个超好用的神器！" \
  douyin_style.wav

Audiobook from Script File (CosyVoice)

python3 scripts/tts.py \
  --platform cosyvoice --voice longxiaoxia_v3 \
  --input chapter1.txt chapter1.wav

Bulk Generation at Lowest Cost (Tencent, inline env vars)

TENCENT_SECRET_ID="xxx" TENCENT_SECRET_KEY="xxx" \
python3 scripts/tts.py \
  --platform tencent --voice 101001 \
  --input course_script.txt course_audio.wav

Premium Quality with Emotion (MiniMax)

MINIMAX_API_KEY="xxx" \
python3 scripts/tts.py \
  --platform minimax --voice female-shaonv \
  "这是一段充满感情的语音合成演示。" premium.wav

Agent-Native CLI Reference

ttscn follows the agent-native-design contract. It serves humans (readable terminal output), AI agents (structured JSON on stdout), and orchestrators (distinct exit codes + idempotency) simultaneously.

JSON Mode

# Explicit JSON envelope mode (independent of --format)
python3 scripts/tts.py --json "你好" out.wav
python3 scripts/tts.py --json --format mp3 "你好" out.mp3

# --format json is a deprecated alias for --json (kept for compatibility)

# Auto-detect: pipe to jq → JSON automatically
python3 scripts/tts.py --list | jq .data.backends[0].name

# Error envelope always structured
python3 scripts/tts.py --json --platform doubao "test" out.wav
# → {"ok":false, "error":{"code":"auth_missing_env","message":"...","retryable":false,...}}

Output Envelope

// Success
{"ok":true, "data":{...}, "meta":{"version":"...","schema_version":"1.2.0","timestamp":"...","ms":123}}

// Error
{"ok":false, "error":{"code":"auth_missing_env","message":"set one of: VOLCENGINE_APPID+VOLCENGINE_ACCESS_TOKEN / VOLCENGINE_API_KEY","retryable":false,"field":"VOLCENGINE_APPID","backend":"doubao"}, "meta":{...}}

Contract: meta.schema_version is a semver string present on every envelope (success and error). The major version bumps only on breaking envelope changes — consumers should assert it matches the schema major they were written against (e.g. video-podcast-maker requires major 1) and fail with a clear error otherwise. An absent schema_version means a pre-contract ttscn release.

Word Boundaries (edge / azure / doubao / minimax / cosyvoice)

For edge, azure, doubao, minimax, and cosyvoice, the success envelope includes native word-level timestamps under data.word_boundaries — absolute seconds within the output file, ascending, 3-decimal rounding. The key is absent for other platforms — and may be absent on doubao/minimax/cosyvoice too when the provider returns no timing payload (minority-language doubao voices; minimax subtitle download failure; cosyvoice-v1 or voices without timestamp support) — so consumers must treat it as optional.

{"ok":true, "data":{
  "output_file": "out.wav",
  "word_boundaries": [
    {"text": "你好", "offset_sec": 0.1,   "duration_sec": 0.45},
    {"text": "世界", "offset_sec": 0.562, "duration_sec": 0.5}
  ]
}}

Use these for subtitle/SRT generation or beat-synced animation without a separate forced-alignment pass.

Exit Codes

Code	Meaning	Agent action
0	Success	Parse data, proceed
1	Internal / runtime error	Report to user, do not retry
2	Validation / fixable error (bad input, missing package)	Fix input or install package, retry allowed
3	Auth / missing credentials	Ask user for API key, do not retry
4	Backend API error	Retry with backoff

Schema Introspection

python3 scripts/tts.py schema backends              # All 14 backends (compact by default)
python3 scripts/tts.py schema backends --full       # All fields (22 per backend)
python3 scripts/tts.py schema backends.doubao       # Single backend full detail
python3 scripts/tts.py schema voices                # All voice presets per backend
python3 scripts/tts.py schema tags                  # Tag definitions
python3 scripts/tts.py schema version               # Version + providers data freshness

# Field filtering for low-token-cost queries
python3 scripts/tts.py schema backends --fields name,cost,supports_clone,supports_ssml

Idempotency

# Orchestrators: retried calls return cached result — no double-billing
python3 scripts/tts.py --idempotency-key "daily-podcast-2026-07-08" --input script.txt out.wav

# Cache at ~/.ttscn_idem/, 7-day TTL, SHA-256 keyed

A cache hit returns the stored result with data.cached: true. If the cached output_file no longer exists on disk, the call re-synthesizes instead of returning a stale success.

Agent Compatibility Flags

# No-ops accepted for agent runtime compatibility (ttscn never prompts)
python3 scripts/tts.py --yes --no-input "text" out.wav