azure-speech-tts（Azure speech TTS 部署）

Azure Speech TTS

Use Azure Speech to turn text or SSML into a local audio file under download/.

What this skill does

• Synthesize plain text into speech
• Synthesize full SSML payloads directly
• Choose voice, output format, rate, pitch, style, and role
• Save the result as a local audio file and print a JSON summary

Configuration

This skill uses a small default config file plus environment variables.

Default config file

File:

• config.json

Default values:

• default_voice: zh-CN-Yunqi:DragonHDOmniLatestNeural
• default_format: mp3
• default_output_dir: download
• default_timeout_seconds: 60

Secret values

Set these in the local shell environment:

• AZURE_SPEECH_KEY
• AZURE_SPEECH_REGION

Optional environment overrides

• AZURE_SPEECH_VOICE
• AZURE_SPEECH_FORMAT

Precedence

Use this order:

1. CLI flag
2. Environment variable
3. config.json
4. Built-in fallback

Quick start

python3 scripts/azure_tts.py \
  --text "你好，这是一段测试语音。" \
  --voice zh-CN-Yunqi:DragonHDOmniLatestNeural \
  --format mp3 \
  --output download/test.mp3

For SSML:

python3 scripts/azure_tts.py \
  --ssml-file temp/input.ssml \
  --format wav \
  --output download/test.wav

Workflow

1. Decide whether the input is plain text or full SSML.
2. Use --text / --text-file for normal narration.
3. Use --ssml / --ssml-file only when the payload already contains a complete <speak> document.
4. Pick the voice and output format, or let config.json supply the defaults.
5. Run scripts/azure_tts.py.
6. Return the generated audio path to the user.

Rules

• Prefer plain text unless the user needs pauses, emphasis, multi-voice content, or expressive styling.
• --ssml input must include a full <speak> root element.
• Default voice is zh-CN-Yunqi:DragonHDOmniLatestNeural if nothing else is set.
• Default output folder is download/.
• If the user does not specify format, use the default MP3 output.
• Do not put secrets in config.json.

Common formats

See references/azure-speech-cheatsheet.md for the format map and examples.

Short aliases supported by the script:

• mp3
• wav
• pcm
• ogg

Useful options

• --voice: Azure voice name, for example en-US-AriaNeural
• --language: SSML xml:lang for plain-text mode
• --rate: speaking rate, for example +10%
• --pitch: pitch adjustment, for example +2st
• --style: expressive style such as cheerful, sad, chat
• --style-degree: strength of the expressive style
• --role: voice role when supported
• --save-ssml: write the generated SSML to a file for inspection
• --dry-run: print the generated SSML without calling Azure

Output

The helper script writes the audio file and prints JSON like:

{
  "ok": true,
  "output_path": "download/test.mp3",
  "format": "audio-24khz-48kbitrate-mono-mp3",
  "voice": "zh-CN-Yunqi:DragonHDOmniLatestNeural",
  "language": "zh-CN",
  "bytes": 123456
}

Use the printed output_path as the deliverable path.