v1.0.1

v1.0.1

• Fixed offline synthesis for every baked core language by loading pinned tokenizer snapshots directly from the image cache, avoiding an unintended Hugging Face metadata request while preserving online installation for optional Spanish and Korean packs.

Full image

docker run --rm -p 8888:8888 --gpus all -v melotts_data:/app/persistent hangrylabs/melotts:v1.0.1@sha256:f0da2ebce41283a5d2dae7125d4776522f95995abdad80773eb2b200fdd1292b

English-family image

docker run --rm -p 8888:8888 --gpus all -v melotts_data:/app/persistent hangrylabs/melotts:v1.0.1_en@sha256:870488904a21cf2e128d80230371270b19dd2a4ea84d70f5505a84afe1bb642f

v1.0.0

v1.0.0

• Corrected the combined project license to AGPL-3.0-only, preserved inherited notices, added network-visible source offers, documented model-specific terms, and pinned audited model revisions.
• Removed Spanish BETO and Korean kykim artifacts from published images; both languages now use explicit terms-aware online installation into a persistent Docker volume and work offline after that first download.
• Added a strict OpenAI-compatible API with model and voice discovery, optional bearer authentication, OpenAI-shaped errors, MP3 defaults, the full 0.25-4.0 speed range, and sentence-streamed MP3/PCM responses.
• Replaced the Gradio interface with the shared Hangry Labs standalone FastAPI UI architecture used by KokoroTTS.
• Added responsive expanded/compact branding, local Lucide icons, WaveSurfer playback and trimming, Generate/Stream/API/System workspaces, model residency controls, and complete demand-driven GPU telemetry with history, hover details, and bounded sampling.
• Added POST /tts/load for loading configured models on demand; existing API behavior and the WAV default remain backward compatible.
• Moved the Docker and package baseline to Python 3.13 and CUDA 13.0 PyTorch wheels.
• Removed Gradio and the obsolete cached-path dependency branch, then regenerated the Python 3.13 lockfile from requirements.in.
• Expanded rapid local iteration tasks to mount the complete melo/, assets/, and VERSION surface without rebuilding the image.
• Made full images and normal local runs default to the six clearly redistributable core model families; English-family images default to EN, EN_V2, and EN_NEWEST.
• Consolidated Docker publication into one strict, metadata-rich pipeline for matching Docker Hub and GHCR full/English-family images.
• Replaced legacy artwork with optimized WebP sets for Melo TTS product surfaces and Hangry Labs organization surfaces.
• Added structural tests for the standalone browser workspace.
• Replaced legacy setup.py packaging with pyproject.toml; wheels now include the runtime VERSION, WebP artwork, standalone UI, and CLI entry points.
• Prevented synthesis text from being written to application logs and serialized inference with model load/purge operations for predictable GPU use.
• Fixed CLI speaker selection for every English model family instead of assuming EN-Default exists.
• Added API, CLI, package-contract, and wheel-content tests plus standard doctor, deps, lint, test, package, and validate Taskfile workflows.
• Regenerated the deployment lock for Linux/Python 3.13 and tightened Docker build context exclusions for tests and private agent files.
• Added optional pitch, tempo, volume, and loudness-normalization controls to the UI, generation API, and sentence-streaming API while keeping neutral defaults backward compatible.
• Added explicit experimental SSML to native generation, streaming, metrics, and the browser UI. The hardened bounded parser supports loaded speaker/model routing, inherited segment prosody, substitutions, character/number/English-ordinal reading, exact breaks, multilingual dialogue, and complete-file normalization while plain text remains the default.
• Added a dedicated SSML examples page with five reproducibly generated MP3 dialogues, exact copyable scripts, branded playback controls, and a validation task for regenerating the media through a live Melo service.
• Standardized public branding on Melo TTS and hangrylabs.app, and removed Tailwind from the examples and 404 pages in favor of purpose-built responsive CSS.
• Documented the distinction between native Melo synthesis controls, post-processing controls, and unsupported named-emotion conditioning.
• Fixed case-sensitive English initialisms and mixed Chinese/English product names: uppercase terms such as US, NLP, LLM, AI, AIGC, and SDXL are now spoken as letters, while lowercase words such as us retain their normal pronunciation and camel-case names retain their word boundaries.
• Fixed reproducible English cleaner errors for dropped-g spellings such as chokin' and jokin', and corrected plugin/plugins to use the short vowel from plug instead of the g2p-en fallback's "ploogin" pronunciation.
• Fixed Chinese BERT/phoneme alignment for verb-一-verb reduplication when Jieba assigns different parts of speech to the repeated verb, preventing duplicate word2ph entries and synthesis failures.
• Reorganized the README around examples, startup, API use, images, development, project context, and release history.

Run this release with either image variant:

Full image

docker run --rm -p 8888:8888 --gpus all -v melotts_data:/app/persistent hangrylabs/melotts:v1.0.0@sha256:e430a9443c5de7479f814ee8c2fdd8a8683e707c19cffc25425d7cce93cb6e3c

English-family image

docker run --rm -p 8888:8888 --gpus all -v melotts_data:/app/persistent hangrylabs/melotts:v1.0.0_en@sha256:15d48aa0e48ca2aeea31cf08aa0872ad10899cdc571dd9891d7c7929ddab0025

v0.1.0

v0.1.0 (11.05.2026)

• Moved the active Docker runtime/build baseline from python:3.10-slim to python:3.11-slim.
• Raised package metadata from python_requires>=3.10 to python_requires>=3.11.
• Refreshed dependency pins for the Python 3.11 line, including newer numpy, pandas, and networkx pins.
• Validated the EN-focused Docker build on Python 3.11 with task imagesmall, python -m pip check, and task localapi.
• Added POST /tts/generate as the preferred synthesis endpoint while keeping legacy POST /tts/convert/tts for backward compatibility.
• Added POST /tts/stream for sentence-level streaming responses plus GET /tts/stream-formats for discovery. The model does not emit token-level audio; streaming starts after each sentence segment is synthesized.

docker run --rm -p 8888:8888 --gpus all -v melotts_data:/app/persistent hangrylabs/melotts:v0.1.0@sha256:a8b9954378dbe3fc871b07a68c3f833d1f5684e92f0c789744fabb1ce08817e0

English-family image

docker run --rm -p 8888:8888 --gpus all -v melotts_data:/app/persistent hangrylabs/melotts:v0.1.0_en@sha256:abf7dfd25fd47121cc06c1a027f6f25a6735d47e7c516324405127496a5c564b