Text to speech

Text-to-Speech Skill

Domain: AI Audio Generation
Version: 4.0.0
Last Updated: 2026-04-15
Author: the AI assistant (your AI assistant)
Source: Patterns from AlexVideos CLI toolkit
Staleness Watch: See EXTERNAL-API-REGISTRY.md for source URLs and recheck cadence

Overview

Cloud-based speech synthesis via Replicate. 15 models spanning MiniMax, Resemble AI, ElevenLabs, Qwen, and Kokoro for narration, audiobooks, voice cloning, and content creation.

---

Model Catalog (15 Models)

Key	Model	Replicate ID	Cost	Cloning	Languages
mm28turbo	Speech 2.8 Turbo	minimax/speech-2.8-turbo	$0.06/1k tokens	❌	40+
mm28hd	Speech 2.8 HD	minimax/speech-2.8-hd	$0.10/1k tokens	❌	40+
mm02turbo	Speech 02 Turbo	minimax/speech-02-turbo	$0.06/1k tokens	❌	40+
mm02hd	Speech 02 HD	minimax/speech-02-hd	$0.10/1k tokens	❌	40+
mm26turbo	Speech 2.6 Turbo	minimax/speech-2.6-turbo	$0.06/1k tokens	❌	40+
mm26hd	Speech 2.6 HD	minimax/speech-2.6-hd	$0.10/1k tokens	❌	40+
mmclone	MiniMax Clone	minimax/voice-cloning	$3/output	✅	—
chatterbox	Chatterbox	resemble-ai/chatterbox	$0.025/1k chars	✅	EN
chatturbo	Chatterbox Turbo	resemble-ai/chatterbox-turbo	$0.025/1k chars	✅	EN
chatpro	Chatterbox Pro	resemble-ai/chatterbox-pro	$0.04/1k chars	✅	EN
chatmlang	Chatterbox Multilingual	resemble-ai/chatterbox-multilingual	variable	✅	Multi
qwentts	Qwen TTS	amphion/qwen3-tts	$0.02/1k chars	✅	10
elevenv3	ElevenLabs v3	elevenlabs/el-multilingual-v3	$0.10/1k chars	❌	Multi
eleventurbo	ElevenLabs Turbo	elevenlabs/el-turbo-v2.5	$0.05/1k chars	❌	Multi
kokoro	Kokoro 82M	jaaari/kokoro-82m	per-second GPU	❌	EN

---

Parameter Support Matrix

Model	text	voice	speed	pitch	volume	emotion	audio ref	language	temperature
mm28turbo	✅	✅	✅	✅	✅	✅	—	✅	—
mm28hd	✅	✅	✅	✅	✅	✅	—	✅	—
mm02turbo	✅	✅	✅	✅	✅	✅	—	✅	—
mm02hd	✅	✅	✅	✅	✅	✅	—	✅	—
mm26turbo	✅	✅	✅	✅	✅	✅	—	✅	—
mm26hd	✅	✅	✅	✅	✅	✅	—	✅	—
mmclone	—	—	—	—	—	—	✅ req	—	—
chatterbox	✅	—	—	—	—	—	✅	—	✅
chatturbo	✅	—	—	—	—	—	✅	—	✅
chatpro	✅	✅	—	✅	—	—	—	—	—
chatmlang	✅	—	—	—	—	—	✅	✅	✅
qwentts	✅	—	—	—	—	—	✅	—	✅
elevenv3	✅	✅	✅	—	—	—	—	✅	—
eleventurbo	✅	✅	✅	—	—	—	—	✅	—
kokoro	✅	✅	✅	—	—	—	—	—	—

---

Model Selection Guide

Scenario	Model	Why
Default / Quick draft	mm28turbo	Fast, cheapest per-token
Studio-grade narration	mm28hd	Highest fidelity, 40+ languages
Clone a specific voice	chatturbo, mmclone	5-second sample, natural pauses
Voice from description	qwentts	No sample needed, describe the voice
Emotion control	mm28turbo/hd	happy, sad, angry, fearful, disgusted, surprised
Non-English content	mm28turbo, elevenv3	Broadest language support
ElevenLabs quality	elevenv3	Premium quality, fine-tuned controls
Lightweight / local	kokoro	Minimal model, fast

---

Voice Presets

MiniMax Speech: Wise_Woman, Deep_Voice_Man, Casual_Guy, Lively_Girl, Young_Knight, Abbess, Childish_Girl, Friendly_Woman, Gentle_Man, Gentle_Woman, Inspirational_girl, Lovely_Girl

Chatterbox Pro: Andy, Luna, Ember, Aurora, Cliff, Josh, William, Orion, Ken

Kokoro: af_heart, af_star, af_sky, am_adam, am_michael, bf_emma, bf_isabella, bm_lewis, bm_george (prefix: af = American female, am = American male, bf = British female, bm = British male)

---

Emotion & Prosody Control (MiniMax)

await replicate.run("minimax/speech-2.8-turbo", {
  input: {
    text: "I am absolutely thrilled with these results!",
    voice: "Lively_Girl",
    emotion: "happy",    // auto, happy, sad, angry, fearful, disgusted, surprised
    speed: 1.2,          // 0.5–2.0 (default 1.0)
    pitch: 5,            // -12 to +12 semitones (default 0)
    volume: 0,           // -6 to +6 dB (default 0)
    language: "en-US",   // 40+ language codes
  },
});

---

Voice Cloning

Chatterbox (5-second sample)

await replicate.run("resemble-ai/chatterbox-turbo", {
  input: {
    text: "Content to speak in the cloned voice",
    audio_prompt: referenceAudioDataURI, // 5+ seconds WAV/MP3
    temperature: 0.7, // 0.1–1.0 (higher = more variation)
  },
});

MiniMax Voice Cloning (Dedicated)

await replicate.run("minimax/voice-cloning", {
  input: {
    audio_sample: referenceAudioDataURI, // High-quality sample
  },
}); // Returns custom voice_id for use in speech models

---

Voice Design (Qwen TTS)

Create a voice from natural language description — no sample needed:

await replicate.run("amphion/qwen3-tts", {
  input: {
    text: "Content to speak",
    tts_mode: "voice_design",
    voice_description: "A warm, friendly female voice with a slight British accent",
    temperature: 0.8,
  },
});

---

ElevenLabs Parameters

await replicate.run("elevenlabs/el-multilingual-v3", {
  input: {
    text: "Content to speak",
    voice_id: "21m00Tcm4TlvDq8ikWAM", // Rachel
    model_id: "eleven_multilingual_v2",
    stability: 0.5,        // 0–1 (higher = more consistent)
    similarity_boost: 0.5, // 0–1 (higher = closer to original voice)
    style: 0.0,            // 0–1 (style exaggeration)
    use_speaker_boost: true,
  },
});

---

macOS Offline Fallback: say

macOS ships 30+ built-in neural voices via the say command. Instant, offline, zero-cost:

say "Hello from the AI assistant"
say -f document.txt
say -o output.m4a --data-format=aac "Dream state finished"
say -v the AI assistant "I am the AI assistant, reading your documentation"

---

Integration with Studio Agents

• Video Generation: Generate narration → merge with avmerge
• Audio Memory: Store voice samples for consistent cloning
• Music Generation: Combine speech with background music

---

Version History

Version	Date	Changes
4.0.0	2026-04-15	Expanded to 15 models from AlexVideos patterns
3.0.0	2026-04-09	Removed Edge TTS, Replicate-only focus
2.5.0	2026-02-09	Speak Prompt command (Edge TTS era)
1.0.0	2026-02-04	Initial implementation via MCP server