What's new.

0.6.4 updates alignment search and introduces the Python SDK.

• Improved long-audio alignment.
• Added the Python SDK.

0.6.3 lets the Alignment page reuse transcript files that already carry word-level timing, translations included.

• Import a JSON with complete word-level timing exported by EdgeSpeak on the Alignment page and the result shows immediately, with no alignment run needed.
• Translations in an EdgeSpeak multi-language JSON are restored on import.

0.6.2 adds downloading audio and video from YouTube, TikTok and other links.

• Paste a video or audio link on the Transcription or Alignment page to download and transcribe or align it, with a download-only option and SRT / VTT subtitle files (a paid feature).
• Site subtitles can be used directly, or switched to on-device transcription with one click.
• The activation page now leads with sign-in through your browser; the license code is the secondary entry.
• The local API, CLI and MCP can design a voice from a one-line description (OmniVoice and FireRedTTS3-Instruct), and CLI and MCP generation can set a translation language.

0.6.1 makes Broadcast, alignment, and transcription steadier on low-spec computers.

• Running several models at once no longer overloads the CPU, making low-spec computers and virtual machines noticeably smoother.

0.6.0 fixes startup and latency issues in voice chat, and officially releases the Lattice-2 alignment model with much higher alignment accuracy.

• Fixed voice chat startup, latency, and related issues, making the whole experience noticeably smoother.
• Officially released the Lattice-2 alignment model, with much higher alignment accuracy.
• Voice chat adds modes such as Buddy companion chat and simultaneous interpretation, making it far more fun to use.
• Added the netease-youdao/Confucius4-R2T2 transcription model, with lower latency.
• Added one-click clearing on the History page.

0.5.6 adds translation and subtitle burn-in: transcripts and alignments can be translated in place.

• Transcription lets you set the audio language, and change it and re-run once you have results.
• Voice chat gained the prompt picker and a Companionship preset, and its cold start is much faster.
• Forced alignment now defaults to Lattice-2, covering 30+ languages.
• Added Translation: transcripts and alignments can be translated segment by segment.
• Added Burn-in: render subtitles into the video itself, with text and image watermarks.

0.5.4 improves conversations, Broadcast, and subtitle editing, with more control over voices and pronunciation and fixes for Windows reliability.

• Improved the Broadcast workspace experience.
• Support for openbmb/VoxCPM2, IndexTeam/IndexTTS-2.5, BreezeBlue/Breeze-TTS-2, and FireRedTTS3 Base, with broader Chinese dialect and multilingual support.
• Fixed language support in Alignment.
• Added a Conversation page for real-time voice chats.
• The CLI supports concurrent requests, with updated code documentation and changelog.
• Updated the Windows version.

0.5.3 updates the macOS desktop app and CLI, with fixes for Broadcast, subtitle alignment, and transcription reliability.

• Improved ASS subtitle export.
• Realtime speech accepts G.711 audio input and reports the actual output sample rate. Session updates avoid restarting recognition when its settings have not changed.
• The CLI includes the updated local runtime and clearer sample rate information in the API and MCP tools.

0.5.2 is a macOS release. Alignment gains a speaker workflow and per-segment proofreading, while both transcription and alignment can now pick an audio track and channel.

• Choose how text is read in Broadcast. For forms such as a year or a price, pin the reading you want.
• Broadcast's script is now an editable workbench: repartition it into segments, regenerate just one segment, edit pronunciations in place, and follow the highlighted text during playback.
• Karaoke captions now include 12 colour presets and a full set of layout controls, including read-text backgrounds and speaker prefixes.
• Broadcast can export at 48 kHz.

0.5.1 is a macOS-only patch release: it fixes incorrect transcription results on some Apple M5 series chips.

0.5.0 adds Alignment: bring a script you already have and export it as subtitles or karaoke captions. Recording now transcribes in real time, and Broadcast gains two top-tier models.

• New Alignment workspace: drop in audio or video plus a script you already wrote, and alignment works out the timestamp for every word — ready to export as karaoke captions.
• Two new Broadcast models: IndexTTS-2.5 clones expressive voices from a reference clip, and FireRedTTS3-Instruct designs a voice from a plain-language instruction.
• Recording now transcribes as you speak, and speaker identification is no longer skipped on recordings.
• Transcript export adds SRT, ASS, and karaoke captions.
• New text normalization endpoint (POST /v1/text/normalizations) and health checks.
• The CLI local server now matches the desktop app's error format and status codes across models, chat, Broadcast, auth, and file transcription.
• CUDA CLI: A100, H100/H20, and data-centre Blackwell now use native cubins instead of relying on PTX JIT.

0.4.4 brings Windows support out of the gate: a slimmer installer with on-demand GPU acceleration, plus new Broadcast models with named voices and style instructions.

• Windows desktop is here as a Preview: local transcription, Broadcast, dictation, and the local API, with NVIDIA GPU acceleration and CPU fallback in one installer.
• The Windows installer is about 70% smaller — roughly 200 MB instead of 730 MB — because GPU acceleration runtimes now download on demand from Settings.
• Broadcast adds three models: a larger cloning model for steadier voices, and two models with 9 built-in official voices each — one also follows style instructions like "furious" or "slow and gentle".
• Import GGUF model repositories from Hugging Face or ModelScope: pick a quantization, see its download size, and keep private-repo tokens in the system credential vault.
• The local API now reports what each Broadcast model can do — named voices, voice cloning, style instructions, streaming — so integrations no longer guess from model names.
• Language model responses over the local API now stream chunk by chunk.
• First transcription engine load on macOS drops from about 12 seconds to about 2.5 seconds.
• The Windows CLI installs with one PowerShell command, and the Linux CLI picks up the same updates.

0.4.1 takes real-time voice conversation from connecting to actually holding a conversation: the first words arrive sooner, interruptions land where you mean them, and the pacing is yours to set.

• The menu bar adds three recording sources: microphone, system audio, or both at once.
• The voice conversation demo lets you adjust pause tolerance, interruption threshold, and gap tolerance mid-call.
• Real-time voice conversation supports role-play personas, ships in seven interface languages, and now comes with integration docs.
• Fixes for voice cloning and voice design on the Broadcast page.

Speaker-aware local transcription, a clearer account activation flow, local language models, and the Realtime API arrive in 0.4.0.

• Local transcription now labels who speaks in each segment; rename the labels and link them to a speaker profile.
• Curated Qwen 3.5 and Gemma 4 local models are now supported.
• The new local Realtime API is in Preview at /v1/realtime over WebSocket for live transcription and voice workflows. It will continue to evolve.
• Account activation is easier to follow: buying a license and activating through an existing account are presented as distinct next steps.
• We have begun preparing the Windows release.

EdgeSpeak learns to speak: this release introduces Broadcast — fully local text-to-speech with voice cloning and voice design.

• Turn text into speech in the new Broadcast workspace, with streaming preview and multi-format export — everything runs on your device.
• Choose among three local models: OmniVoice by default, plus Qwen3-TTS models for voice cloning and voice design, downloaded in-app on first use.
• Clone your own voice from a short recording, or design a brand-new voice from a text description; rename or delete voices anytime.
• Use Broadcast everywhere: the OpenAI-compatible /v1/audio/speech endpoint on the local API, the CLI, and MCP.
• Long texts are synthesized in segments, so you can listen to each part as soon as it is ready.
• Full offline mode for buyout licenses: activate once online, then transcribe, align, and broadcast without a network.
• The Linux CLI installer now picks the GPU acceleration package matching your NVIDIA card, with a safe CPU fallback.

A smoother desktop release with batch file transcription, faster local model loading, and steadier progress feedback.

• Import several audio or video files at once; EdgeSpeak keeps the selected order and gives a clear message when the batch is too large.
• Local model loading is faster, so transcription starts with less waiting.
• Longer transcription jobs show steadier progress and return useful results sooner.
• Microphone permission status is more reliable, and starting a new recording no longer carries over stale batch selections.
• The sidebar can now be resized and stays cleaner in narrow windows.

A reliability update for local transcription, app permissions, and the local gateway.

• Local file transcription now releases post-processing resources more promptly after jobs finish.
• Microphone permission prompts appear only after an explicit recording or permission action.
• The local gateway, CLI, and MCP runtime are more reliable when starting and binding local ports.

A small update that makes local model setup and app updates feel smoother.

• Missing local models now show clearer setup and retry states.
• macOS updates are more reliable after download and app relaunch.
• The sidebar update button is shorter, avoiding overlap with the app name and version.

A developer-tools release: EdgeSpeak's local engine now speaks CLI, MCP, and a wider local API, with fully tunable semantic segmentation.

• Run transcription, forced alignment, and text segmentation from the terminal with the bundled edgespeak-cli.
• Let AI assistants use EdgeSpeak locally through the built-in MCP server.
• Call new local API endpoints for forced alignment and text segmentation; transcription now returns word-level timestamps.
• Tune semantic segmentation in Settings — segment length and boundary margins — shared as the default by the app, API, and CLI.
• Transcribe with EdgeSpeak Lite, a lighter and faster default local model, and follow post-processing progress in real time.

A dictionary release: teach EdgeSpeak your terms with hotwords, correction rules, and encrypted multi-device sync.

• Add names and domain terms to the dictionary so they are recognized correctly.
• Create correction rules that automatically rewrite frequently misheard words.
• Sync your dictionary across devices with encryption.
• Read release notes in the update dialog and adjust the app font size.

A transcription workbench release focused on editing, playback accuracy, and macOS compatibility.

• Reset the whole workspace when you want to start fresh.
• Edit transcript text and segment timing with millisecond precision.
• Play audio using the exact segment start and end timestamps.
• Use the notarized macOS build on Apple Silicon with macOS 14.0 or later.

Desktop release focused on clearer license activation, better upgrade guidance, and a more dependable update path.

Desktop update with a more native macOS feel, clearer local transcription states, improved segment playback, and stronger install/update reliability.

Initial early access release for local audio and video transcription, desktop license activation, and the local speech gateway.