MiniMax Voice Maker

Professional text-to-speech skill with emotion detection, voice cloning, and audio processing capabilities powered by MiniMax Voice API and FFmpeg.

Capabilities

File structure:

mmVoice_Maker/
├── SKILL.md                       # This overview
├── mmvoice.py                     # CLI tool (recommended for Agents)
├── check_environment.py           # Environment verification
├── requirements.txt
├── scripts/                       # Entry: scripts/__init__.py
│   ├── utils.py                   # Config, data classes
│   ├── sync_tts.py                # HTTP/WebSocket TTS
│   ├── async_tts.py               # Long text TTS
│   ├── segment_tts.py             # Segment-based TTS (multi-voice, multi-emotion)
│   ├── voice_clone.py             # Voice cloning
│   ├── voice_design.py            # Voice design
│   ├── voice_management.py        # List/delete voices
│   └── audio_processing.py        # FFmpeg audio tools
└── reference/                     # Load as needed
    ├── cli-guide.md               # CLI usage guide
    ├── getting-started.md         # Setup and quick test
    ├── tts-guide.md               # Sync/async TTS workflows
    ├── voice-guide.md             # Clone/design/manage
    ├── audio-guide.md             # Audio processing
    ├── script-examples.md         # Runnable code snippets
    ├── troubleshooting.md         # Common issues
    ├── api_documentation.md       # Complete API reference
    └── voice_catalog.md           # Voice selection guide

Main Workflow Guideline (Text to Speech)

6-step workflow: [step1]. Verify environment

[step2-preparation]⚠️NOTE: Before processing the text, you must read voice-catalog.md for voice selection.

[step2]. Process text into script → <cwd>/audio/segments.json. Note: [Step2.4] is really important, you must check it twice before sending the script to the user.

[step2.5]. ⚠️ Generate preview for user confirmation (highly recommended for multi-voice content)

[step3]. Present plan to user for confirmation

[step4]. Validate segments.json

[step5]. Generate and merge audio → intermediate files in <cwd>/audio/tmp/, final output in <cwd>/audio/output.mp3

[step6]. ⚠️ CRITICAL: User confirms audio quality FIRST → THEN cleanup temp files (only after user is satisfied)

<cwd> is Claude's current working directory (not the skill directory). Audio files are saved relative to where Claude is running commands.

Step 1: Verify environment

python check_environment.py

Checks:

• Python 3.8+
• Required packages (requests, websockets)
• FFmpeg installation
• MINIMAX_VOICE_API_KEY environment variable

If API key is not set, ask user for keys and set it:

export MINIMAX_VOICE_API_KEY="your-api-key-here"

Step 2: Decision and Pre-processing

⚠️ MOST IMPORTANT PRINCIPLE: Gender Matching First

Before selecting voices, you MUST always match gender first. This is non-negotiable.

Golden Rule:

If a character is male → use male voice If a character is female → use female voice If a character is neutral/other → choose appropriate neutral voice

Why this matters:

• Violating gender matching (e.g., male character with female voice) breaks immersion
• Even if personality traits match, gender comes first
• This is especially critical for classic literature, historical content, and professional narration

Examples:

Decision guide: Evaluate based on:

• Does the user specify a model? → Use that model, or use the default one "speech-2.8"
• Is multi-voice needed? → Different voice_id per speaker/character
• For speech-2.8: emotion is auto-matched (leave emotion empty)
• For older models: manually specify emotion tags

Use case scenarios:

Processing Workflow (4 sub-steps):

Step 2.1: Text Segmentation and Role Analysis First, segment your text into logical units and identify the role/character for each segment.

Key principle (Important!): Split by logical unit, NOT simply by sentence

When to split (Important!):

• Different speakers clearly marked
• Narrator vs. character dialogue (in fiction/audiobooks/interview etc.)
• In some scenarios (like audiobooks, multi-voice fiction etc.), where speaker's identity is important, split when narration and dialogue mix in the same sentence.

When NOT to split (Important!):

• Third-person narration like "John said..." or "The reporter noted..."
• Quoted speech in narration (in documentary/podcast/report etc.) should keep in narrator's voice
• Keep in narrator's voice unless specific characterization is needed

Decision depends on use case:

Example1: Single Voice (speech-2.8) For single-voice content (e.g., news, announcements, articles), segment only by length while maintaining the same voice:

[
  {"text": "First part of the article (under 1,000,000 chars)...", "role": "narrator", "voice_id": "female-shaonv", "emotion": ""},
  {"text": "Second part of the article (under 1,000,000 chars)...", "role": "narrator", "voice_id": "female-shaonv", "emotion": ""},
  {"text": "Third part of the article (under 1,000,000 chars)...", "role": "narrator", "voice_id": "female-shaonv", "emotion": ""}
]

Example2: Audiobook with characters (speech-2.8) In audiobooks (multi-voice fiction), split when narration and dialogue mix in the same sentence:

[
  {"text": "The detective entered the room.", "role": "narrator", "voice_id": "", "emotion": ""},
  {"text": "\"Who's there?\"", "role": "female_character", "voice_id": "", "emotion": ""},
  {"text": "she whispered.", "role": "narrator", "voice_id": "", "emotion": ""},
  {"text": "\"It's me,\"", "role": "male_character", "voice_id": "", "emotion": ""},
  {"text": "he replied calmly.", "role": "narrator", "voice_id": "", "emotion": ""}
]

Example3: Documentary/podcast narration (speech-2.8) Quoted speech in narration stays in narrator's voice (no need to split):

[
  {
    "text": "The scientist explained, \"The results show significant improvement in all test groups.\"",
    "role": "narrator",
    "voice_id": "",
    "emotion": ""
  },
  {
    "text": "According to the latest report, the economy has grown by 3% this quarter.",
    "role": "narrator",
    "voice_id": "",
    "emotion": ""
  }
]

**Note:** In the preliminary `segments.json`:
- Fill in the `text` field with segment content
- Fill in the `role` field to identify the character (narrator, male_character, female_character, host, guest, etc.)
- Leave `voice_id` empty (to be filled in Step 2.2)
- Leave `emotion` empty for speech-2.8 models


**Step 2.2: Voice Selection**

After segmenting and labeling roles, analyze all detected characters in your text. Consult [voice_catalog.md](reference/voice_catalog.md) **Section 1 "How to Choose a Voice"** to match voices to characters.

**⚠️ CRITICAL: Follow the two-step selection process below**

**Path A — Professional domains (Story/Narration, News/Announcements, Documentary):**
If the content belongs to one of these three professional domains, prioritize selecting from the recommended voices in **voice_catalog.md Section 2.1** (filter by scenario + gender). These voices are specifically optimized for their professional use cases.

**Path B — All other scenarios:**
Select from **voice_catalog.md Section 2.2**, following this strict priority hierarchy:

1. **First: Match Gender** (non-negotiable) — Male characters MUST use male voices, female characters MUST use female voices
2. **Second: Match Language** — The voice MUST match the content language (Chinese content → Chinese voice, Korean content → Korean voice, English content → English voice, etc.). Never assign a voice from the wrong language.
3. **Third: Match Age** — Determine the age group (Children / Youth / Adult / Elderly / Professional) and select from the corresponding subsection in Section 2.2
4. **Fourth: Match Personality & Role** — Choose the best fit based on personality traits, tone, and character role

**Voice Selection Decision Tree:**

Is this a professional domain (Story/News/Documentary)? ├── YES → Select from voice_catalog Section 2.1 (filter by scenario + gender) └── NO → Select from voice_catalog Section 2.2: Step 1: Match Gender ├── Male character → Male voices only └── Female character → Female voices only Step 2: Match Age Group └── Children / Youth / Adult / Elderly / Professional Step 3: Match Language └── Filter to voices matching the content language Step 4: Match Personality & Role └── Choose best fit by tone, personality, character role

**Step 2.3: Emotions Segmentation** *(For non-2.8 series models only)*
For models other than speech-2.8 series, analyze emotions in your segments:
- For **long segments**, split further based on **emotional transitions**
- Add appropriate **emotion tags** to each segment
- Refer to Section 3 in [text-processing.md](reference/text-processing.md) for emotion tags and examples
- Skip this step for speech-2.8 models (emotion is auto-matched)

**Emotion Tags:**
- For speech-2.6 series (speech-2.6-hd and speech-2.6-turbo): happy, sad, angry, fearful, disgusted, surprised, calm, fluent, whisper
- For older models: happy, sad, angry, fearful, disgusted, surprised, calm (7 emotions)


**Step 2.4: Check and Post-processing**
Finally, review and optimize your script:
- Verify segment length limits (async TTS ≤1,000,000 characters)
- Clean up conversational text (remove speaker names if needed)
- Ensure consistency in voice and emotion tags
- **Critical check for multi-voice content**: For audiobooks, multi-voice fiction, or content where dialogue is presented from a first-person perspective, verify that narration and dialogue mixed in the same sentence are properly split.

  **When splitting IS needed (first-person dialogue in fiction/audiobooks):**
  
  Example: `"John asked, 'Where are you going?'"` should be split into:
  - Segment 1: `"John asked, "` - uses narrator voice (describes who is speaking)
  - Segment 2: `"Where are you going?"` - uses the character's voice (actual dialogue in first-person)

  This ensures proper voice differentiation: descriptive narration uses the narrator's voice, while the character's spoken words use the character's designated voice.

  **When splitting is NOT needed (third-person quotes in podcast/documentary/news):**
  
  In podcasts, documentaries, or news reports, quoted speech is typically presented in third-person narrative style - the speaker's words are being reported, not performed. Keep these as one segment with the narrator's voice and remove the speaker's name at the beginning:
  
  - `"Welcome to our show." → narrator voice, remove the speaker's name (like "The host said:") at the beginning
  - `"According to experts, 'This technology represents a significant breakthrough.'" → keep as one segment (narrator voice)
  - `"Scientists noted, 'The experimental results exceeded our expectations.'" → keep as one segment (narrator voice)
- **If the split is missing**: Go back to Step 2.1 and ensure dialogue portions are separated from narration with appropriate role labels.

**Create segments.json:**
After completing all 4 sub-steps, save the final `segments.json` to `<cwd>/audio/segments.json`.


### Step 2.5: Generate Preview for User Confirmation (Highly Recommended)

**For multi-voice content (audiobooks, dramas, etc.), always generate a preview first.**

This saves time and prevents waste when voice selections need adjustment.

**How to generate a preview:**
1. Create a smaller segments file with 10-20 representative segments (include all characters)
2. Generate the preview audio
3. Ask user to listen and confirm voice choices

**Preview segments.json example:**
```json
[
  {"text": "Narration opening...", "role": "narrator", "voice_id": "...", "emotion": ""},
  {"text": "Male character speaks...", "role": "male_character", "voice_id": "...", "emotion": ""},
  {"text": "Female character speaks...", "role": "female_character", "voice_id": "...", "emotion": ""},
  {"text": "More dialogue...", "role": "...", "voice_id": "...", "emotion": ""}
]

Preview command:

python mmvoice.py generate segments_preview.json -o preview.mp3

When user confirms preview:

• Use the same voice selections for the full segments.json
• No need to re-select voices

---

Step 3: Present plan to user for confirmation

Before proceeding to validation and generation, present the segmentation plan to the user and wait for confirmation:

Present to the user:

• Roles identified: List all characters/speakers in the text
• Voice assignments: Show which voice_id is assigned to each role (include voice characteristics from voice_catalog.md)
• Model being used: Explain why this model was selected
• Language: Confirm the primary language of the content
• Emotion approach: Auto-matched (speech-2.8) or manual tags (older models)

Example confirmation message:

I've analyzed the text and created a segmentation plan:

**Roles and Voices:**
- Narrator: male-qn-jingying (deep, authoritative, suitable for storytelling)
- Protagonist: female-shaonv (bright, energetic, youthful)
- Antagonist: male-qn-qingse (cool, menacing)

**Model:** speech-2.8-hd (recommended - automatic emotion matching)
**Language:** Chinese
**Segments:** 8 segments total

Please review and confirm:
1. ⚠️ **Gender Verification**: Do the voice genders match the character genders?
   - [Narrator: Male ✓] [Protagonist: Female ✓] [Antagonist: Male ✓]
2. ⚠️ **Language Verification**: Do the voice languages match the content language?
   - [All voices: Chinese ✓]
3. Are the voice assignments appropriate for each character (age, personality)?
4. Should any segments be combined or split differently?
5. Any other changes you'd like to make?

**After generation:**
- I'll generate a preview first for you to review
- Only after you confirm the audio quality will I clean up temporary files
- If not satisfied, I'll re-generate and we iterate until you're happy

Reply "confirm" to proceed, or let me know what to adjust.

Wait for user response:

• If user confirms → Proceed to Step 4 (validate)
• If user suggests changes → Update segments.json and present the plan again for confirmation

Step 4: Validate segments.json (model, emotion, voice_id validation)

Before generating audio, validate the segments file:

# Default: speech-2.8-hd (auto emotion matching)
python mmvoice.py validate <cwd>/audio/segments.json

# Specify model for context-specific validation
python mmvoice.py validate <cwd>/audio/segments.json --model speech-2.6-hd

# Validate voice_ids against available voices (slower, requires API call)
python mmvoice.py validate <cwd>/audio/segments.json --validate-voices

# Combined options (recommended)
python mmvoice.py validate <cwd>/audio/segments.json --model speech-2.6-hd --validate-voices

# Use `--verbose` to see segment details
python mmvoice.py validate <cwd>/audio/segments.json --model speech-2.6-hd --validate-voices --verbose

Emotion Validation checks:

Voice ID validation: With --validate-voices:

• Calls API once to get all available voices
• Validates each voice_id against the list
• Shows errors for invalid voice_ids (blocks validation)

Step 5: Generate and merge audio

Generate audio for all segments and merge into final output.

File placement (default behavior if user doesn't specify):

<cwd>/                      # Claude's current working directory
└── audio/                  # Created automatically
    ├── tmp/                # Intermediate segment files
    │   ├── segment_0000.mp3
    │   ├── segment_0001.mp3
    │   └── ...
    └── <custom_audio_name>.mp3             # Final merged audio, name can be customized

Where <cwd> is Claude's current working directory (where commands are executed).

• If -o is not specified, output goes to <cwd>/audio/output.mp3
• Intermediate files go to <cwd>/audio/tmp/
• After user confirms the final audio, ask whether to delete <cwd>/audio/tmp/

Basic usage:

# Default: speech-2.8-hd, output to <cwd>/audio/output.mp3
python mmvoice.py generate <cwd>/audio/segments.json

# Specify output path
python mmvoice.py generate <cwd>/audio/segments.json -o <cwd>/audio/<custom_audio_name>.mp3

# Specify model if needed
python mmvoice.py generate <cwd>/audio/segments.json --model speech-2.6-hd

Skip existing segments (for rate limit retries):

# Only generate segments that don't exist yet - skips already-generated files
python mmvoice.py generate <cwd>/audio/segments.json --skip-existing

Error handling:

• If a segment fails, the script reports which segment and why
• Use --continue-on-error to generate remaining segments despite failures
• Use --skip-existing to skip already successfully generated segments (recommended for retries after rate limit)
• The script automatically uses fallback merging if FFmpeg filter_complex fails

Step 6: Confirm and cleanup

⚠️ CRITICAL: Never delete temp files until user confirms!

After generation completes, you MUST follow this exact sequence:

Step 6.1: Report generation result to user

✓ Audio saved to: <output_path>
  Generated: X/Y segments
  Intermediate files in: <cwd>/audio/tmp/

Step 6.2: Ask user to confirm audio quality Ask the user to listen to the audio and confirm:

1. Is the audio quality satisfactory?
2. Are all voices appropriate?
3. Any adjustments needed?

Step 6.3: Wait for user response

Step 6.4: Only after user confirms, offer cleanup

After confirming audio quality, temporary files can be deleted with:
rm -rf <cwd>/audio/tmp/

NEVER execute rm -rf on temp files without explicit user confirmation!

If user is NOT satisfied:

• Do NOT delete temp files
• Discuss what needs to be adjusted
• Re-generate affected segments if needed
• Ask for confirmation again

Other Usage

Use the following when the task involves voice creation, single-voice TTS (sync/async), or audio processing instead of the main segment-based workflow. Each subsection gives CLI commands, script paths, and the reference doc to open for details.

Voice creation (clone / design / list)

• Purpose: Create custom voices from audio (clone) or from a text description (design); list system and custom voices.
• CLI (entry point: mmvoice.py):

  python mmvoice.py clone AUDIO_FILE --voice-id VOICE_ID   # Clone from 10s–5min audio
  python mmvoice.py design "DESCRIPTION" --voice-id ID      # Design from text
  python mmvoice.py list-voices                             # List all voices
  

• Scripts: scripts/voice_clone.py (clone), scripts/voice_design.py (design), scripts/voice_management.py (list/manage).
• Documentation: reference/voice-guide.md — cloning (quick + high-quality + step-by-step), design workflow, management.

Text-to-speech (sync / async)

• Purpose: Single-voice TTS: sync for short text (≤10k chars), async for long text (up to 1M chars); optional streaming.
• CLI:

  python mmvoice.py tts "TEXT" -o OUTPUT.mp3 [-v VOICE_ID] [--model MODEL]
  

• Scripts: scripts/sync_tts.py (HTTP/WebSocket sync), scripts/async_tts.py (async task + poll).
• Documentation: reference/tts-guide.md — sync TTS, async TTS, streaming, segment-based production.

Audio processing (merge / convert / normalize)

• Purpose: Merge files (with optional crossfade), convert format, normalize loudness, trim.
• CLI:

  python mmvoice.py merge FILE1 [FILE2 ...] -o OUTPUT [--crossfade MS]
  python mmvoice.py convert INPUT -o OUTPUT [--format FORMAT]
  

• Script: scripts/audio_processing.py (merge, convert, normalize, trim).
• Documentation: reference/audio-guide.md — format conversion, merging (filter_complex + concat demuxer fallback), normalization, trimming, optimization.

Segment-based TTS (main workflow)

• CLI: validate and generate as in Steps 4–5 above.
• Script: scripts/segment_tts.py.
• Documentation: reference/cli-guide.md, reference/api_documentation.md.

---

Reference documents (on-demand)

Open these when you need concrete usage, parameters, or troubleshooting. Paths are relative to the skill root.

Important notes

Requirements

• Python: 3.8 or higher
• API Key: MINIMAX_VOICE_API_KEY environment variable must be set
• FFmpeg: Required for audio processing (merge, convert, normalize)
  • Install: brew install ffmpeg (macOS) or sudo apt install ffmpeg (Ubuntu)

Limits and constraints

• Text length: Sync TTS ≤10,000 chars; async TTS ≤1,000,000 chars
• Voice cloning: Audio must be 10s–5min duration, ≤20MB, formats: mp3/wav/m4a
• Voice expiration: Custom voices (cloned/designed) expire after 7 days if not used with TTS

Special features

• Pause insertion: Use <#x#> in text where x = pause duration in seconds (0.01–99.99)
  • Example: "Hello<#1.5#>world" creates 1.5s pause between words
• Supported emotions: happy, sad, angry, fearful, disgusted, surprised, calm, fluent, whisper
  • speech-2.8: automatic matching; speech-2.6: all 9; older models: first 7

Troubleshooting

• Run python check_environment.py to diagnose setup issues
• See troubleshooting.md for common problems and solutions
• Check getting-started.md for detailed setup instructions