Voice Clone Skill

A self-initializing, zero-configuration voice cloning skill. It manages a background TTS daemon that keeps heavy model weights in memory for fast inference. Supports multiple engines and unlimited text length.

Quick reference

Item	Value
Entry script	INLINECODE0
Output

Single line: absolute path to generated .ogg file | | Attachment format | MEDIA:<output_path> | | Default engine | F5-TTS (env TTS_BACKEND=f5) | | Host/Port config | .env (TTS_SERVER_HOST, TTS_SERVER_PORT) |

When to use this skill

• - The user sends a voice memo or audio file and you need to reply with audio.
• The user says "read this aloud", "speak to me", "use my voice", "voice mode".
• The conversation context implies a spoken reply is expected.
• The user provides a reference audio and asks you to mimic their voice.

Step-by-step usage

1. Identify inputs

You need two things:

• - ref_audio: The absolute local path to the user's reference audio file (the voice to clone). This is typically the audio file the user just sent, saved by the ASR system (e.g., openai-whisper).
• text: The text content you want to speak. Generate this as you normally would — think of your reply, then voice it.

2. Run the synthesis

Execute this command:

CODEBLOCK0

Optional parameters:

• - --speed 1.2 — Speak faster. Range: 0.5 to 2.0. Default: 1.0.
• INLINECODE10 — Save the generated audio file to a specific absolute folder path. Default: server/generated_audio/.

Example with all options:
CODEBLOCK1

3. Handle the output

The script prints a single absolute path on stdout (e.g., /path/to/reply_a1b2c3d4.ogg).
Append it to your response using the attachment format:

CODEBLOCK2

4. Important constraints

• - Do NOT manually start python app.py or manage the backend. The run_tts.sh script auto-detects, auto-installs, and auto-starts everything.
• First run is slow (~30-60 seconds) because it downloads model weights and loads them into memory. Subsequent calls are fast.
• Long texts work automatically. The engine splits text into sentences, synthesizes each chunk, and stitches them seamlessly. No length limit.

Controlling voice characteristics

Speed (all engines)

The --speed parameter adjusts speaking rate:

Value	Effect
INLINECODE16	Slow, deliberate, suitable for elderly listeners
INLINECODE17

Natural conversational speed (default) |
| 1.3 | Brisk, suitable for news or briefings |
| 1.5+ | Fast, compressed delivery |

F5-TTS supports speed natively. Other engines use ffmpeg post-processing (atempo filter), which gives good results but may slightly affect quality at extreme values.

Emotion and tone

These models use acoustic feature extraction from the reference audio — they do not accept text-based emotion tags like [happy] or [sad].

The emotion of the output is determined entirely by the reference audio.

To control emotion, select or prepare reference audio that carries the desired tone:

Desired tone	Reference audio strategy
Calm, neutral	Use a reference clip where the speaker talks normally
Excited, happy

Use a reference clip where the speaker sounds enthusiastic |
| Angry, intense | Use a reference clip with raised voice and sharp intonation |
| Sad, melancholic | Use a reference clip with slow, downcast delivery |
| Whispering | Use a reference clip where the speaker whispers |

Practical approach for Agents: If the user has sent multiple voice messages, choose the one whose emotional tone best matches the context of your reply. If only one reference is available, use it as-is — the model will approximate the speaker's general style.

ChatTTS Specifics: This engine supports inline emotion tags in text: [laugh], [uv_break] (pause). It also supports voice cloning when a reference audio is provided.

Available engines

Engine	ID	Install	Size	Clone	Speed support	Best for
F5-TTS	INLINECODE24	INLINECODE25	~1.5GB	✅	Native	Highest quality cloning
CosyVoice

cosyvoice | bash scripts/install_cosyvoice.sh | ~1.5GB | ✅ | ffmpeg | Natural Chinese prosody | | ChatTTS | chattts | bash scripts/install_chattts.sh | ~400MB | ✅ | ffmpeg | Dialogue with emotion tags | | OpenVoice | openvoice | bash scripts/install_openvoice.sh | ~300MB | ✅ | ffmpeg | Ultra fast, tiny footprint |

Switch engines by setting the environment variable before the server starts:
CODEBLOCK3

Uninstalling

CODEBLOCK4

File structure

CODEBLOCK5

References

• - Read references/architecture.md for system architecture and design rationale.