elevenlabs-smart-tts 0.1.4

Smart TTS library with ElevenLabs and OpenRouter text enhancement

elevenlabs-smart-tts

High-level Python library for expressive text-to-speech with ElevenLabs and LLM-powered text enhancement via OpenRouter.

Pass raw text plus task context (language, style, emotion, use case) — the library picks a voice, enriches the text with Eleven v3 audio tags, and returns synthesized audio.

Features

• SmartTTS facade — one pipeline from text to audio
• Voice caching — local diskcache catalog with offline list_voices() / get_voice()
• Automatic voice selection — by voice_id, description, use case, style, and language
• LLM text enhancement — audio tags, punctuation, and normalization via OpenRouter
• Eleven v3 first — expressive tags like [whispers], [excited], [short pause]
• Typed errors & retries — resilient HTTP clients for ElevenLabs and OpenRouter

Installation

pip install elevenlabs-smart-tts

Or from source:

git clone https://github.com/vpuhoff/elevenlabs-smart-tts.git
cd elevenlabs-smart-tts
uv sync --dev

Quick start

1. Copy .env.example to .env and fill in your API keys:

cp .env.example .env

2. Run synthesis:

from pathlib import Path

from elevenlabs_smart_tts import SmartTTS, SynthesisTask

tts = SmartTTS.from_env()
tts.sync_voices()

result = tts.synthesize_to_file(
    SynthesisTask(
        text="Welcome to our customer support service.",
        language="en",
        style="professional",
        emotion="warm",
        use_case="conversational",
        voice_description="warm professional conversational",
    ),
    Path("output.mp3"),
)

print(result.enhanced_text)

See example.py for a full runnable example.

Task parameters

SynthesisTask accepts free-text hints that guide voice selection and LLM text enhancement. After sync_voices(), inspect your cached catalog:

for voice in tts.list_voices():
    print(voice.name, voice.labels.get("use_case"), voice.description)

The examples below come from the ElevenLabs premade voice catalog.

use_case

Used for voice matching against the ElevenLabs voice label labels.use_case (exact match scores highest).

Value	Typical voices
conversational	Casual, agentic, podcast-style voices (e.g. Roger, Eric, Juniper)
informative_educational	Clear educators, broadcasters (e.g. Alice, Matilda, Daniel)
narrative_story	Storytellers, audiobook voices (e.g. George, Daria Reels)
advertisement	Promo and ad reads (e.g. Bill)
social_media	Short-form, trendy content
characters_animation	Character and animation voices
entertainment_tv	TV and entertainment narration

customer_support is not an ElevenLabs label — it still helps the LLM, but for voice selection prefer conversational or pass voice_description="professional support warm".

# Support-style call center message
SynthesisTask(
    text="Thanks for calling. How can I help you today?",
    language="en",
    use_case="conversational",
    style="professional",
    emotion="warm",
    voice_description="trustworthy professional",
)

# Audiobook / long-form narration
SynthesisTask(
    text="Chapter one. It was a dark and stormy night.",
    use_case="narrative_story",
    style="warm",
    emotion="calm",
)

# E-learning explainer
SynthesisTask(
    text="Today we'll learn how photosynthesis works.",
    use_case="informative_educational",
    style="professional",
    emotion="neutral",
)

style

Free-form delivery hint. Affects the LLM enhancement prompt and weak voice matching against voice name, description, and custom tags.

Common values that match premade voice descriptions:

Value	Effect
professional	Formal, clear delivery
casual / conversational	Relaxed, everyday tone
warm	Friendly, inviting tone
neutral	Balanced, informative
dramatic	Strong emphasis, expressive pacing
playful	Light, energetic tone
sympathetic	Soft, empathetic delivery

SynthesisTask(text="...", style="professional")   # business / IVR
SynthesisTask(text="...", style="casual")         # laid-back chat
SynthesisTask(text="...", style="dramatic")       # emotional scene

emotion

Free-form mood hint for LLM text enhancement only (drives audio tags like [excited], [whispers], [sighs]). Does not filter voices.

Value	Typical audio tag behavior
warm	Friendly, reassuring tone
calm	Steady, subdued delivery
excited	Higher energy, [excited] tags
sympathetic	Soft, caring tone
curious	Questioning, engaged tone
appalled / sarcastic	Strong expressive tags
neutral	Minimal emotional markup

SynthesisTask(text="...", emotion="warm")         # customer greeting
SynthesisTask(text="...", emotion="excited")      # product launch
SynthesisTask(text="...", emotion="sympathetic")  # apology or support
SynthesisTask(text="...", emotion="neutral")     # plain narration

Combining parameters

Scenario	Example values
Customer support (EN)	use_case="conversational", style="professional", emotion="warm"
News / podcast intro	use_case="informative_educational", style="neutral", emotion="calm"
Audiobook chapter	use_case="narrative_story", style="warm", emotion="calm"
Social reel	use_case="social_media", style="playful", emotion="excited"
Ad read	use_case="advertisement", style="confident", emotion="excited"

Configuration

Required environment variables

Variable	Description
ELEVENLABS_API_KEY	ElevenLabs API key
OPENROUTER_API_KEY	OpenRouter API key
OPENROUTER_API_TTS_PROMPT_MODEL	LLM for text enhancement (e.g. anthropic/claude-3.5-sonnet)

Optional environment variables

Variable	Default	Description
ELEVENLABS_CACHE_DIR	~/.cache/elevenlabs-smart-tts	Local cache directory
ELEVENLABS_DEFAULT_MODEL	eleven_v3	Default TTS model
ELEVENLABS_DEFAULT_OUTPUT_FORMAT	mp3_44100_128	Audio output format
ELEVENLABS_DEFAULT_VOICE_ID	—	Fallback voice when auto-selection fails
OPENROUTER_BASE_URL	https://openrouter.ai/api/v1	OpenRouter API base URL

Programmatic configuration is also supported:

from elevenlabs_smart_tts import SmartTTS, SmartTTSConfig, TTSModel

config = SmartTTSConfig(
    elevenlabs_api_key="...",
    openrouter_api_key="...",
    openrouter_tts_prompt_model="anthropic/claude-3.5-sonnet",
    default_model=TTSModel.ELEVEN_V3,
)
tts = SmartTTS(config)

Usage

Synthesis pipeline

from elevenlabs_smart_tts import SmartTTS, SynthesisTask, TTSModel

tts = SmartTTS.from_env()
tts.sync_voices()

result = tts.synthesize(
    SynthesisTask(
        text="Are you serious? I can't believe you did that!",
        voice_id="your-voice-id",
        model=TTSModel.ELEVEN_V3,
        style="dramatic",
        emotion="appalled",
    )
)

audio_bytes = result.audio
enhanced_text = result.enhanced_text

Preview enhanced text without TTS

enhanced = tts.enhance_text_only(
    SynthesisTask(
        text="Thanks for calling. How can I help?",
        language="en",
        style="sympathetic",
    )
)

One-liner

from elevenlabs_smart_tts import synthesize

result = synthesize(
    "Hello world",
    language="en",
    style="neutral",
)

Async API

import asyncio
from pathlib import Path

from elevenlabs_smart_tts import AsyncSmartTTS, SynthesisTask, asynthesize

async def main() -> None:
    async with AsyncSmartTTS.from_env() as tts:
        await tts.sync_voices()
        result = await tts.synthesize_to_file(
            SynthesisTask(text="Hello world", language="en"),
            Path("output.mp3"),
        )
        print(result.enhanced_text)

asyncio.run(main())

# Or as a one-liner:
result = asyncio.run(asynthesize("Hello world", language="en"))

Voice management

voices = tts.list_voices(language="en", tags=["narration"])
voice = tts.get_voice("voice-id")

tts.sync_voices(force=True)  # refresh cache from ElevenLabs API

Supported TTS models

Model	Best for
eleven_v3	Expressive speech, audio tags, emotions
eleven_multilingual_v2	Multilingual, high voice similarity
eleven_flash_v2_5	Low latency, conversational agents

Development

uv sync --dev
uv run pytest
uv run ruff check .

License

MIT — see LICENSE.

Links

• GitHub repository
• PyPI package
• Design specification (Russian)