screencastgen

Qwen Backend

Local Qwen3-TTS backend used directly by the CLI or behind the inference server.

Source: screencastgen/providers/tts/qwen_backend.py

Overview

Implements the TTSBackend protocol using Qwen3-TTS for local text-to-speech synthesis. It supports:

multi-language generation
optional voice cloning with ref_audio_path and ref_text
batched synthesis through synthesize_batch(...) for the inference server

Class: `QwenTTS`

Properties

Property	Value	Description
`max_chunk_bytes`	`1500`	Conservative text limit exposed to chunking and validation
`output_format`	`"wav"`	Output audio format

Constructor

QwenTTS(
    model_name=None,
    ref_audio_path=None,
    ref_text=None,
    language="en-US",
    device="auto",
)

Param	Type	Default	Description
`model_name`	`str \\| None`	`None`	Model alias or full Hugging Face model name
`ref_audio_path`	`str \\| None`	`None`	Reference audio file for voice cloning
`ref_text`	`str \\| None`	`None`	Transcript for the reference audio
`language`	`str`	`"en-US"`	Language code mapped to Qwen language labels
`device`	`str`	`"auto"`	Compute device

Methods

Method	Description
`synthesize(text, output_path)`	Generate one audio clip and write it to disk
`synthesize_batch(texts, language=None, ref_audio_path=None, ref_text=None)`	Generate multiple WAV payloads in one model call

Models

Default model:

Qwen/Qwen3-TTS-12Hz-0.6B-Base

Built-in aliases:

0.6b
0.6B
1.7b
1.7B

Supported language map includes:

English
Chinese
Japanese
Korean
German
French
Russian
Portuguese
Spanish
Italian

BackendSpec

Registered in TTS Registry with:

Contexts: cli, server
Capabilities: local, voice_clone
Extra args: --model

The inference server currently expects this backend to provide synthesize_batch(...).

The model download hook uses huggingface_hub.snapshot_download() to populate the cache without constructing the Qwen model. If the selected model value resolves to a local path, the download step leaves it unchanged.

Dependencies

Qwen Backend
├── qwen-tts            (deferred import)
├── torch               (deferred import)
├── soundfile           (deferred import)
├── TTS Base        (resolve_device, BackendSpec)
└──▶ registered in TTS Registry