Qwen3-TTS Studio

The Hangry Labs Studio is the default browser UI in the root Docker image.

Interface languages

The interface supports English (/en), Norwegian Bokmål (/nb), Polish (/pl), Japanese (/ja), Simplified Chinese (/zh) and Spanish (/es). English is the default. The header picker updates the interface without reloading, preserving unsaved dialogue, reference recordings, takes, output audio and operator edits. The preference stays in browser local storage. Opening / restores it; an explicit language URL takes precedence. Missing translations fall back to English.

Interface language does not change the synthesis language. Qwen’s ten speech languages remain the same; Norwegian and Polish are interface translations only. Example buttons use the selected speech language, or English when Auto is selected. Changing interface language never translates user text, voice names, saved scripts or protocol/model identifiers. Language switching is paused during generation. The Studio combines an audio editor, recording controls and GPU monitoring with Qwen’s generation workflows.

task init
task start
# Open http://127.0.0.1:7860

Generate loads the selected model when needed; the runtime badge shows its state. Tiny downloads uncached models. Full seeds its baked files into an empty product cache before starting the UI and defaults to offline operation. See the container workflow for build targets and publishing rules. The five model revisions are fixed to the versions listed in the licensing guide. System selects the default model for each workflow and whether it stays in GPU memory. By default, 0.6B Base, 0.6B CustomVoice and 1.7B VoiceDesign load at startup. Kept models can coexist; other models are released after the complete request. A global toggle disables keeping weights. These settings survive browser and container replacement.

Workflows

View	Model	Workflow
Generate	CustomVoice, Base and document-defined VoiceDesign	Magic dialogue, text, SSML and SSML-H; native or saved voices; locked takes and saved scripts.
Voice design	1.7B VoiceDesign	Describe a voice, generate a sample, then use or save it as a cloning reference.
Voice clone	0.6B or 1.7B Base	Upload, record, or choose a saved reference. Clone speaker identity alone or use an exact reference transcript, with optional delivery direction.
Voices	Saved audio-backed voices	Search metadata, preview references, add portraits, use voices and restore their creation workflow for refinement.
API	Same inference engine	Inspect capabilities and endpoint documentation.
System	Operator controls	Load/unload models, configure ASR, and inspect demand-driven GPU telemetry.

All voice workflows offer complete WAV output and incremental playback. Qwen supports Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, and Italian, plus automatic language detection. Model capabilities drive the controls; unsupported combinations are rejected before loading.

Save a generated recording or an uploaded reference explicitly. The save dialog accepts a description, up to twelve tags and an optional portrait. Search the Voices tab by name, description, tags or language. Use voice selects the saved reference in Generate; Edit details changes metadata and portraits without replacing audio.

Refine restores the original creation workflow, model, text, direction, sampling controls and recorded seed when those settings were saved. Clones also restore their source reference and transcript. Original and processed recordings remain separate, with saved finishing settings restored alongside them. Older voices without a recipe open as cloning references.

Generate fresh audio to enable replacement of the voice being refined. Saving offers the original or a current processed preview as the cloning reference; stale processed previews cannot be selected. Replacing a voice requires the replacement toggle and confirmation. Leaving it unchecked saves a separate voice under a unique name. Finish refining ends the replacement context.

Magic character promotion includes tags, an optional portrait, the generated turn transcript and creation recipe. Existing names require explicit replacement; failed saves leave the script character intact. Saved voices are audio-backed Base cloning references even when their original creation used VoiceDesign or CustomVoice. Restored settings aid repeatability, but model/runtime changes can change the resulting audio.

Generate uses a full-width left settings rail, central authoring surface and saved-script rail. Its toolbar holds actual Qwen model controls, seed and output processing. Magic turns support voice/language selection, direction, prosody, pauses, substitutions, preview, regeneration, lock, download and reordering. Switching between plain text and Magic preserves the text without adding dialogue. Keep individual Magic turns within 400 characters; longer text can be synthesized through the text/SSML modes with automatic splitting.

Save a named script to keep its document, settings, takes and current output in the product volume. Imports/exports use SSML-H. Script folders, tags and search organize the library. Saving remains explicit; ordinary editing and synthesis do not automatically persist prompts or recordings. SSML-H scope="profile" is an explicit instruction to persist a designed voice after successful generation. See SSML-H for exact semantics and API for endpoint contracts.

Generation settings include temperature, top K, repetition penalty, chunk size, streaming decoder context (1–200 frames, default 25), maximum audio length, optional seed, and whole-text/incremental-text delivery. Text delivery controls model input; streaming controls browser playback. These are independent settings. The audio editor supports playback, volume, speed, seeking, trimming, download, and sharing where supported by the browser.

Voice clone offers an optional Delivery direction field, such as “Speak slowly and gently, with a reassuring tone.” It is passed to the Base engine in both complete and streaming generation. Choose Speaker and reference transcript and supply the exact reference words for the recommended instruction path. Base instruction following is experimental and can vary; identity-only cloning may ignore the direction. This does not carry the same instruction-control qualification as 1.7B CustomVoice. Leave it blank for ordinary voice cloning. API clients use the existing instruct option alongside mode: "clone" and xvec_only: false; capability metadata marks Base instruction support experimental.

Cloning defaults to Speaker and reference transcript with incremental text input. Enter the exact words or use Transcribe reference. Select Speaker identity only when a transcript is unavailable. Generate and Voice design default to whole-text input. Explicit text-delivery choices stay with their workflow while the page is open. Reset controls restores the current workflow’s default.

The recording surface lets you select a microphone or use the system default. Refresh the list after connecting hardware; browser permission reveals device names. Device changes and permission errors are reported beside recording. Audio chunk controls accept every integer from 1 to 64 codec frames and show approximate audio duration. Playback streaming is independent of text input.

Generate’s Stream mode previews MP3 progressively, including within long-text and dialogue units. Boundary compaction retains up to 600 ms at internal edges; speed/prosody and whole-stream normalization add their own filter buffering. Browser playback therefore starts later than model/SSE first-audio timings.

The finishing menu includes Signalsmith Voice, FFmpeg Voice Shape and advanced mastering controls. Processed WAV previews preserve float samples at the native 24 kHz rate. Normalization is opt-in and original audio remains available.

Protect long audio splits Generate text into units of at most 400 characters. The Maximum audio length selection applies to each model call, not to the complete joined recording. A request supports up to 6000 spoken characters, 64 speech units and five minutes of final audio. Dedicated Design/Clone tabs use a single call with up to 1000 characters. See the API streaming and long-text details for current limits and the independent-chunk behavior.

Metrics report audio length, first-audio latency, generation time, and audio-seconds per generation-second. First-audio and generation timings begin after model loading; reference preparation and audio decoding are included. The API reports total inference time including queue wait and model loading. Generation diagnostics distinguish queue wait, model loading, reference preparation and cache reuse, streaming model TTFA/generation, browser first-chunk receipt, scheduled playback start and live playback buffer. Model timings cover token prefill/decode when reported; unavailable measurements are not estimated. Breakdowns overlap and should not be summed. Scheduled playback does not measure speaker hardware latency. Waiting requests show queue position and can be stopped.

Prepared voice prompts are reused by audio content, transcript and cloning settings within the loaded model. An LRU cache retains at most eight entries and 8 MiB of tensor data in system RAM. No raw recording is retained by this cache; uploaded temporary files are deleted after inference. Unloading or switching the model clears prepared references. Explicitly saved voices remain in the product volume.

Tab changes retain current work; reloading clears prompts and generated audio. Save voice explicitly stores a reference and transcript in the product volume. That library and ASR settings survive browser/container restarts. No inherited reference recordings are bundled or fetched.

Only clone authorized speakers and process recordings you may use. Do not use cloned voices for deceptive impersonation. A model license does not grant rights to somebody else’s identity or copyrighted input.

Reference transcription

Connect an existing OpenAI-compatible service such as Qwen3-ASR-STT. Save or test its URL in System. No ASR server runs inside this product, and no local ASR container is launched for testing. A service on another machine leaves local VRAM for TTS.

QWEN_STUDIO_ASR_URL supplies the initial connection when no saved setting exists. task start forwards this environment variable; saved System settings take precedence. Leave the URL blank to disconnect. Reference audio is sent to the chosen service only when Transcribe reference is pressed. Generation and manual transcripts work without ASR.

Ports and API

The UI and Studio API share port 7860, bound to loopback by default. For another host interface, run task start BIND_ADDRESS=0.0.0.0 PORT=7860. This local Studio has no account/login system. Microphone capture requires localhost or HTTPS and browser permission. Its UI assets, fonts, and waveform editor do not require a CDN.

Local API documentation is at /docs, with its schema at /openapi.json. /studio/capabilities describes the models, controls and limits. /studio/generate returns base64 WAV plus metrics; /studio/generate/stream returns server-sent events carrying WAV chunks. It also reports queue/loading progress. With ?streaming=false, the same progress route surrounds complete inference and returns one complete audio chunk; the ordinary /studio/generate route retains its JSON response. Both accept multipart data containing an options JSON field and an optional reference audio file. Separate /v1/audio/speech and native /tts/* routes provide the product API. Independent compact/advanced MCP endpoints can be enabled in System and return expiring audio links. Both default to disabled.

Validation

The supported runtime is NVIDIA/Torch with local Qwen inference source. The Studio includes model selection, all three voice modes, upload/recording, transcription, cloning methods, sampling controls and downloadable output.

task validate
task test-runtime
task smoke
task smoke -- --all-models

Smoke checks generate real GPU samples and transcribe them with the selected ASR service. Audio/results are written under ignored artifacts/studio-smoke/. Set QWEN_STUDIO_ASR_URL in your shell or an untracked .env file before running Task, or pass ASR_URL=https://asr.example.com to task smoke. Direct script invocations accept --asr-url. The smoke test requires an explicit connection; it does not fall back to a local ASR service. Generated samples are sent to that service for transcription. These short correctness checks are not representative quality/speed benchmarks. See development tasks and licensing for the runtime locking and image-release qualification requirements.