TypeWhisper/typewhisper-mac v1.8.2

on GitHub

TypeWhisper 1.8.2 is an update to the stable 1.8 macOS release. It makes ElevenLabs dictations much faster, stops cloud engines from uploading audio again for the live preview, lets you import your own Parakeet models, adds word timestamps to the HTTP API and fixes dictation into Flutter apps and the kitty terminal.

Highlights

• ElevenLabs 1.1.0: realtime dictations in Auto mode no longer wait about 15 seconds after the transcript arrives. In our tests a 3.5-second dictation took about half a second instead of 16 seconds. The plugin also adds the Scribe v2 Medical model, more languages and word timestamps.
• Live preview: with a cloud engine that cannot stream, such as Microsoft AI, Groq or ElevenLabs, the live preview no longer uploads the last 10 seconds of audio again every 3 seconds. It runs on a loaded local engine instead, while the final text still comes from your dictation engine.
• Parakeet 1.7.0: Import Model… adds your own Parakeet TDT models converted to Core ML, for example a fine-tune for one language.
• HTTP API: verbose_json responses can include the time of each word.

Dictation

• With Live preview engine set to Match dictation engine, the preview now runs on a loaded local engine that streams your language, such as Apple Speech, while the final text still comes from your dictation engine. Without such an engine no preview is shown, unless live text in the text field or the Live Transcript window is on; these keep using the cloud engine then. To keep the previous behavior, choose your cloud engine under Settings > Appearance > Live preview engine. Before, the preview sent about three times the audio of each dictation to the cloud engine.
• Dictating into Flutter apps for macOS shows the text in the field. TypeWhisper wrote it into Flutter's hidden input view, which reported success while the visible field stayed empty; it now pastes there.
• Dictating into the kitty terminal no longer ends with "No text field detected" when kitty shows the text a moment late. kitty passes typed text on to the shell, which echoes it back, so a field that has not changed yet does not mean the text went nowhere. TypeWhisper now counts the dictation as inserted there, as it already does for web inputs.
• Apple Speech offers Apple's regional Chinese models separately, so you can pick Hong Kong Cantonese with Traditional Chinese characters. Before, Chinese could fall back to the model of another region.
• The "New in TypeWhisper" window has a single Continue button and stays in front of Settings while TypeWhisper is active.

Plugins, API and CLI

• ElevenLabs 1.1.0: realtime dictations in Auto mode end as soon as the transcript arrives; the plugin used to wait for ElevenLabs to close the finished session. Auto mode still uses the batch API for Scribe v2 Medical, with active dictionary terms and for recordings over 30 seconds, so those take as long as before. If a realtime session sends no final transcript, the batch API transcribes the whole recording instead of returning partial text.
  • Transcriptions through the batch API keep Scribe's word timestamps. verbose_json API responses get segments split at sentence ends and pauses instead of an empty segments list, and word timestamps when requested. Speaker detection aligns ElevenLabs transcripts by word as well. Short realtime dictations report word timestamps and, with automatic language detection, the detected language.
  • The new Scribe v2 Medical model, tuned for clinical audio at the Scribe v2 price, transcribes in batch. A saved selection of the deprecated Scribe v1 moves to Scribe v2.
  • The detected language is reported with the two-letter code TypeWhisper uses (de instead of deu), so German and other language-specific formatting applies with automatic detection.
  • Javanese, Asturian, Cebuano, Irish, Igbo, Kabuverdianu, Kurdish, Kyrgyz, Ganda, Luo, Northern Sotho, Odia, Umbundu, Wolof, Xhosa and Zulu can be selected; Hawaiian and Volapük, which ElevenLabs rejects, are no longer offered.
  • Realtime errors whose type does not say "error", such as rate_limited, quota_exceeded or invalid_request, switch to the batch API at once instead of after an eight-second wait. When an API key's own credit quota is used up, the error shows ElevenLabs' message instead of "Invalid API key".
• Parakeet 1.7.0: Import Model… in the plugin settings adds your own Parakeet TDT models converted to Core ML, from a folder or a Hugging Face repository. Raw .mlpackage conversions are compiled during the import. Fine-tunes that add tokens to the vocabulary, such as the Greek Parakeet v3 fine-tune, transcribe correctly. Imported models appear in the model list and can be deleted there.
• Language Detection 1.0.1 tells Hong Kong Cantonese (zh-HK) apart from Mandarin. Before, all Chinese variants counted as one language, so dictation always used the first Chinese language you selected.
• HTTP API: verbose_json responses can include the time of each word. Send timestamp_granularities[]=word (as in OpenAI's API) to get a words array with word, start and end in seconds; /v1/transcribe/local-file takes "timestamp_granularities": ["word"]. Responses without the parameter stay unchanged. Engines that report word times, such as Parakeet and ElevenLabs, fill it.
• CLI: typewhisper completions bash|zsh|fish prints a shell completion script. It completes commands and options, file paths, and engine and model IDs from the running app. The CLI is part of the direct download edition.

Compatibility and distribution

• Requires macOS 14.0 or later and supports Apple silicon and Intel Macs.
• Every direct download release includes the iCloud bridge. The Mac App Store edition ships 1.8.2 from the same commit.
• Parakeet 1.7.0 requires TypeWhisper 1.8.2. ElevenLabs 1.1.0 runs on TypeWhisper 1.8.0 and later; word timestamps through the HTTP API need 1.8.2. Language Detection 1.0.1 requires TypeWhisper 1.8.1.
• Built with Xcode 27.1.