Audio Generation¶

Tarash Gateway provides a unified interface for text-to-speech (TTS) and speech-to-speech (STS) generation across ElevenLabs, Cartesia, Fal.ai, Sarvam, and Hume — using a single generate_tts or generate_sts call regardless of provider.

Pass an AudioGenerationConfig and a TTSRequest to generate_tts(), or an STSRequest to generate_sts().

Text to Speech (TTS)¶

Basic TTS¶

Pass text and a voice ID to generate speech audio.

from tarash.tarash_gateway import generate_tts
from tarash.tarash_gateway.models import AudioGenerationConfig, TTSRequest

config = AudioGenerationConfig(
    provider="elevenlabs",
    model="eleven_multilingual_v2",
    api_key="YOUR_ELEVENLABS_KEY",
)
request = TTSRequest(
    text="Welcome to Tarash, the unified AI media gateway.",
    voice_id="JBFqnCBsd6RMkjVDRZzb",
)
response = generate_tts(config, request)
print(response.content_type)  # e.g. "audio/mpeg"

ElevenLabs models →

With Output Format¶

Control the audio format, sample rate, and bitrate via output_format.

from tarash.tarash_gateway import generate_tts
from tarash.tarash_gateway.models import AudioGenerationConfig, TTSRequest, AudioOutputFormat

config = AudioGenerationConfig(
    provider="cartesia",
    model="sonic-3",
    api_key="YOUR_CARTESIA_KEY",
)
request = TTSRequest(
    text="High quality audio with custom format settings.",
    voice_id="a0e99841-438c-4a64-b679-ae501e7d6091",
    output_format=AudioOutputFormat(format="wav", sample_rate=44100),
)
response = generate_tts(config, request)
print(response.content_type)  # "audio/wav"

Cartesia models →

With Voice Settings¶

Some providers accept voice_settings to fine-tune the generated speech.

from tarash.tarash_gateway import generate_tts
from tarash.tarash_gateway.models import AudioGenerationConfig, TTSRequest

config = AudioGenerationConfig(
    provider="elevenlabs",
    model="eleven_multilingual_v2",
    api_key="YOUR_ELEVENLABS_KEY",
)
request = TTSRequest(
    text="This voice is tuned for maximum expressiveness.",
    voice_id="JBFqnCBsd6RMkjVDRZzb",
    voice_settings={
        "stability": 0.3,
        "similarity_boost": 0.8,
        "style": 0.7,
        "use_speaker_boost": True,
    },
)
response = generate_tts(config, request)

ElevenLabs voice settings →

Multilingual TTS¶

Providers like Sarvam specialize in specific languages. Pass language_code to select the target language.

from tarash.tarash_gateway import generate_tts
from tarash.tarash_gateway.models import AudioGenerationConfig, TTSRequest

config = AudioGenerationConfig(
    provider="sarvam",
    model="bulbul:v3",
    api_key="YOUR_SARVAM_KEY",
)
request = TTSRequest(
    text="नमस्ते, तरश गेटवे में आपका स्वागत है।",
    language_code="hi-IN",
)
response = generate_tts(config, request)

Sarvam models →

Expressive TTS¶

Hume's Octave models support expressive speech with emotion and speed control.

from tarash.tarash_gateway import generate_tts
from tarash.tarash_gateway.models import AudioGenerationConfig, TTSRequest

config = AudioGenerationConfig(
    provider="hume",
    model="octave-2",
    api_key="YOUR_HUME_KEY",
)
request = TTSRequest(
    text="I'm so excited to tell you about this!",
    voice_id="Kora",
    voice_settings={
        "description": "excited and enthusiastic",
        "speed": 1.2,
    },
)
response = generate_tts(config, request)

Hume models →

TTS via Fal.ai¶

Fal.ai hosts MiniMax Speech and Qwen 3 TTS models.

from tarash.tarash_gateway import generate_tts
from tarash.tarash_gateway.models import AudioGenerationConfig, TTSRequest

config = AudioGenerationConfig(
    provider="fal",
    model="fal-ai/minimax/speech-2.8-hd",
    api_key="YOUR_FAL_KEY",
)
request = TTSRequest(
    text="Hello from MiniMax Speech on Fal.",
    voice_id="male-qn-qingse",
)
response = generate_tts(config, request)

MiniMax Speech → · Qwen 3 TTS →

Speech to Speech (STS)¶

STS takes an existing audio clip and re-synthesizes it in a different voice, preserving the original speech content. Pass an STSRequest with the source audio.

Basic STS¶

from tarash.tarash_gateway import generate_sts
from tarash.tarash_gateway.models import AudioGenerationConfig, STSRequest

config = AudioGenerationConfig(
    provider="elevenlabs",
    model="eleven_multilingual_v2",
    api_key="YOUR_ELEVENLABS_KEY",
)
request = STSRequest(
    audio="https://example.com/source-speech.mp3",
    voice_id="JBFqnCBsd6RMkjVDRZzb",
)
response = generate_sts(config, request)
print(response.content_type)

ElevenLabs STS →

Voice Changer with Cartesia¶

Cartesia's STS uses its voice changer API to transform audio into a target voice.

from tarash.tarash_gateway import generate_sts
from tarash.tarash_gateway.models import AudioGenerationConfig, STSRequest, AudioOutputFormat

config = AudioGenerationConfig(
    provider="cartesia",
    model="sonic-3",
    api_key="YOUR_CARTESIA_KEY",
)
request = STSRequest(
    audio="https://example.com/source-speech.wav",
    voice_id="a0e99841-438c-4a64-b679-ae501e7d6091",
    output_format=AudioOutputFormat(format="wav", sample_rate=44100),
)
response = generate_sts(config, request)

Cartesia STS →

Async Generation¶

Every function has an async variant. Use generate_tts_async or generate_sts_async for non-blocking workflows.

from tarash.tarash_gateway import generate_tts_async
from tarash.tarash_gateway.models import AudioGenerationConfig, TTSRequest

config = AudioGenerationConfig(
    provider="elevenlabs",
    model="eleven_flash_v2_5",
    api_key="YOUR_ELEVENLABS_KEY",
)
request = TTSRequest(
    text="Fast async generation with ElevenLabs Flash.",
    voice_id="JBFqnCBsd6RMkjVDRZzb",
)
response = await generate_tts_async(config, request)

Provider Comparison¶

Provider	TTS	STS	Highlights
ElevenLabs	✅	✅	29 languages, voice cloning, fine-grained voice settings
Cartesia	✅	✅	Low-latency Sonic models, voice changer API
Fal.ai	✅	—	MiniMax Speech (interjections, pauses), Qwen 3 TTS
Sarvam	✅	—	11 Indian languages, Bulbul models
Hume	✅	—	Expressive Octave models with emotion control