Qwen 3 TTS (via Fal.ai)¶

Qwen 3 text-to-speech model hosted on Fal.ai, with voice customization and fine-grained sampling control.

Quick Example¶

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

config = AudioGenerationConfig(
    provider="fal",
    model="fal-ai/qwen-3-tts",
    api_key="YOUR_FAL_KEY",
)

request = TTSRequest(
    text="Hello! This is a test of the Qwen 3 TTS model.",
    voice_id="Chelsie",
    language_code="en",
)

response = generate_tts(config, request)
print(f"Audio size: {len(response.audio)} chars (base64)")

Parameters¶

Parameter	TTSRequest field	Required	Notes
Text	`text`	✅	Text to synthesize
Voice	`voice_id`	—	Voice selection
Style prompt	`prompt`	—	Voice style guidance
Language	`language_code`	—	Language / locale

Sampling Parameters¶

Pass via extra_params for fine-grained control over generation:

Parameter	Type	Notes
`temperature`	`float`	Sampling temperature
`top_k`	`int`	Top-k sampling filter
`top_p`	`float`	Nucleus (top-p) sampling
`repetition_penalty`	`float`	Penalty for repeated tokens
`max_new_tokens`	`int`	Maximum token count

Sub-Talker Parameters¶

Control the sub-talker (additional speaker modeling) via extra_params:

Parameter	Type	Notes
`subtalker_dosample`	`bool`	Whether to use sampling
`subtalker_top_k`	`int`	Sub-talker top-k
`subtalker_top_p`	`float`	Sub-talker nucleus sampling
`subtalker_temperature`	`float`	Sub-talker temperature

request = TTSRequest(
    text="Hello world",
    voice_id="Chelsie",
    extra_params={
        "temperature": 0.7,
        "top_p": 0.9,
        "max_new_tokens": 2048,
    },
)