Inworld AI

Text-to-Speech platform with voice cloning, audio markups, and timestamp alignment.

Quick Navigation

When to Use

• Text-to-speech audio generation
• Voice cloning from 5-15 seconds of audio
• Emotion-controlled speech ([happy], [sad], etc.)
• Word/phoneme timestamps for lip sync
• Custom pronunciation with IPA

Models

Minimal Example

import requests, base64, os

response = requests.post(
    "https://api.inworld.ai/tts/v1/voice",
    headers={"Authorization": f"Basic {os.getenv('INWORLD_API_KEY')}"},
    json={"text": "Hello!", "voiceId": "Ashley", "modelId": "inworld-tts-1.5-max"}
)
audio = base64.b64decode(response.json()['audioContent'])

Key Features

• 15 languages — en, zh, ja, ko, ru, it, es, pt, fr, de, pl, nl, hi, he, ar
• Instant cloning — 5-15 seconds audio, no training
• Audio markups — [happy], [laughing], [sigh] (English only)
• Timestamps — word, phoneme, viseme timing for lip sync
• Streaming — /voice:stream endpoint
• TTS-2 steering — natural-language bracketed directions such as [say excitedly] or [whisper in a hushed style]
• Delivery mode — STABLE, BALANCED, CREATIVE trade consistency for emotional range
• Cross-lingual synthesis — reuse one voice across multiple languages; voice localization improves native-sounding output

Release Highlights (TTS-2)

• Realtime TTS-2 becomes the new primary model line via modelId="inworld-tts-2".
• Steering moves beyond the older fixed emotion tags: free-form bracketed directions can control style, pitch, speed, intensity, and non-verbals.
• Multilingual coverage expands with production quality across 15 languages and broader experimental coverage beyond that.
• deliveryMode adds a stability-vs-creativity knob, and specifying language matters more for cross-lingual output quality.

Release Updates (August 2026)

• New inworld-tts-2-flash model: lowest latency and cost, full language coverage, instant voice cloning, and timestamp alignment. Steering and Professional Voice Cloning remain exclusive to inworld-tts-2.
• Steering instructions now persist until explicitly changed: a reserved [reset] tag ends a styled passage, and a <break/> pause no longer clears the active instruction. A request-level instruction field on Synthesize Speech is an alternative to inline tags.
• Cross-lingual voice synthesis and voice localization improve native-sounding output when one voice is reused across languages.

Prohibitions

• Audio markups work only in English
• Use ONE emotion markup at text beginning
• Match voice language to text language
• Instant cloning may not work for children's voices or unique accents

Links

• Documentation
• Changelog
• Platform