Faster Qwen3-TTS Studio

Derived from Faster Qwen3-TTS by Andres Marafioti.
Copyright (c) 2026 Andres Marafioti. Original wrapper code is MIT licensed.
Modifications and distribution work are maintained by Hangry Labs.

This distribution also contains original and adapted Apache-2.0 code from:

  Qwen3-TTS, Copyright 2026 The Alibaba Qwen team; and, where indicated
  upstream, The Qwen team, Alibaba Group and the HuggingFace Inc. team.
  https://github.com/QwenLM/Qwen3-TTS

  Transformers 5 compatibility changes by Andres Marafioti, incorporated from
  https://github.com/andimarafioti/Qwen3-TTS/tree/distribution/qwen-tts-hf
  including QwenLM/Qwen3-TTS pull request #360.
  The Qwen model and 12 Hz inference code are maintained locally; Hangry Labs
  omits the upstream demo and 25 Hz codec and adapts tokenizer exports/loading.
  Exact incorporated revisions are recorded in the packaged upstreams.toml and
  docs/UPSTREAMS.md. The separately published qwen-tts-hf wheel is not bundled.

  Qwen3-TTS-streaming, derived from Qwen3-TTS and maintained by dffdeeq
  and contributors.
  https://github.com/dffdeeq/Qwen3-TTS-streaming

The local talker input construction and streaming adaptation differ from
upstream to support static caches, CUDA Graphs, chunked audio, and wrapper
integration. Later Hangry Labs changes are recorded in source history.

The source distribution therefore has MIT and Apache-2.0 obligations. The
MIT text in LICENSE does not relicense Apache-2.0 portions, model assets,
dependencies, or native libraries. Retain this notice and the applicable
license texts. See THIRD_PARTY_NOTICES.md for the consolidated notices and texts.

The project acknowledges nano-qwen3tts-vllm by Thanh Sang for CUDA Graph
inspiration. That project is MIT licensed; inspiration is not a claim that
the project is a required runtime dependency.

No endorsement by upstream authors, model providers, or NVIDIA is implied.

The Hangry Labs Studio shell, example gallery, media components, styles, GPU
monitor, Magic take bundle, dialogue script storage and adapted API contracts are
adapted from Hangry Labs VoxCPMTTS (Apache-2.0):
https://github.com/hangry-labs/VoxCPMTTS
Qwen-specific workflow, runtime, and API changes are maintained in this project.
Sentence planning, dialogue boundary compaction and mastering filters also
adapt VoxCPMTTS implementations for Qwen's codec budgets and native sample rate.
Browser library notices accompany their vendored assets and are summarized in
THIRD_PARTY_NOTICES.md.

The manual benchmark measurements, transcript normalization, dialogue corpus,
queue comparison methodology and report structure are also adapted from
Hangry Labs VoxCPMTTS (Apache-2.0). Qwen checkpoint routing, isolated workers,
native comparison adapter and related benchmark changes are maintained here.

MCP transport, temporary audio artifacts, bounded reference fetching and release
transition tooling also adapt Hangry Labs VoxCPMTTS patterns (Apache-2.0).

The runtime uses the MCP Python SDK (MIT), Copyright 2024 Anthropic, PBC.
Its original license remains installed and is included in runtime inventories.

The runtime uses SSML-H Tools, Copyright 2026 Hangry Labs, Apache-2.0.
It implements parts of https://github.com/hangry-labs/ssml-h-spec.
SSML-H is based on W3C SSML 1.1 but is not a W3C Recommendation and is not
endorsed by W3C. The package's original NOTICE and license remain installed.