Crisp ASR

Crisp ASR is a desktop-only Obsidian plugin by letschips that connects directly to Doubao Speech Recognition.

Version 0.4.4

Crisp ASR keeps Doubao Speech Recognition as the dedicated transcription engine and adds an optional, separate AI text-processing layer.

Manual smart processing

• Finish a file or live transcription first, then explicitly choose 润色整理, 重点提炼, or 自定义.
• Nothing is sent to a text model until the user clicks one of those actions.
• 润色整理 corrects punctuation, paragraphs, obvious ASR errors, filler words, and mechanical repetition without removing meaningful information.
• 重点提炼 removes rambling and repetition, then organizes key facts, conclusions, decisions, action items, owners, dates, and open questions.
• 自定义 runs the saved Prompt with support for {{transcript}}, {{title}}, {{date}}, {{audio_file}}, and {{duration}}.

Provider choice

• Choose 火山方舟, OpenAI, Claude, DeepSeek, or a custom OpenAI-compatible endpoint.
• Ark, OpenAI, DeepSeek, and custom endpoints use the Chat Completions contract. Claude uses the native Anthropic Messages contract.
• The text-model Secret and model name are independent from the Doubao Speech Recognition Secret.
• Provider presets keep their official Base URL. The custom provider accepts a user-supplied Base URL.
• A short connection test validates the selected Secret, model, and endpoint.

Preview and raw-transcript protection

• Processing opens a side-by-side preview with the untouched transcript and generated result.
• Users can regenerate, copy, discard, or explicitly write the result.
• The default write mode inserts a marked ## 智能整理 section immediately before the raw ASR section.
• Re-running smart processing replaces only the previous generated section.
• If the raw transcript changes while generation is running, Crisp ASR refuses the write and asks the user to regenerate.
• An optional output mode creates a separate linked smart-result note instead.
• Long transcripts are split at readable boundaries and processed in order; multi-part key extraction receives a final consolidation pass.
• Cancelling stops the workflow between provider calls and never writes a partial result.

Data boundary

• Audio files, microphone input, and explicitly authorized computer audio are sent only to Volcengine for speech recognition.
• Transcript text is sent only to the AI provider selected by the user and only after a manual smart-processing action.
• Both Speech and AI API keys are stored through Obsidian SecretStorage rather than plugin data.json.
• Crisp ASR does not operate a proxy server.

Version 0.3.2

• Add About Crisp ASR to the bottom of the settings page with the plugin's core purpose and the linked author attribution.

• Update the input meter and elapsed time without rebuilding the full sidebar, and preserve transcript scroll position when new text arrives.

• Keep every persistent job reachable through an expandable task list.

• Retry Volcengine 55xxxxxx temporary service failures even when the HTTP response itself is successful.

• Cancel an in-progress live startup cleanly when the plugin unloads.

• Clarify no-note output behavior and use the full control width for computer-only capture.

• Move a legacy Crisp ASR plaintext accessToken into Obsidian SecretStorage and preserve its live resource ID during the first upgraded load.

• Choose microphone, computer audio, or computer audio plus microphone for live transcription.

• Use the macOS sharing picker for every computer-audio session. Crisp ASR never starts system capture silently and never reads or stores shared video.

• Select a preferred microphone and monitor the mixed input level in the sidebar.

• Optionally preserve the exact mixed ASR input as WebM/Opus under Crisp ASR/Audio; this remains disabled by default.

• Embed a saved live recording in the resulting Markdown transcript.

• Persist file-transcription jobs across restarts and run them serially.

• Retry temporary network, timeout, rate-limit, and service failures twice. Permanent authentication, missing-file, decode, and size failures remain in the panel for manual retry or removal.

• Resume interrupted queued or running jobs after Obsidian reloads.

• Use neutral real-time dictation wording for meetings, interviews, voice notes, classes, and other recording contexts.

• Avoid nested backdrop-blur compositor layers in Obsidian's translucent macOS window, which caused expanding shadows and screen trails during live updates.

• Show a compact floating live strip with elapsed time, transcript preview, panel access, and a safe stop-and-write action.

• Test the configured API key from settings or the command palette with a 100 ms silent probe.

• Transcribe an audio attachment within two lines of the editor cursor.

• Control the panel and live session through obsidian://crisp-asr?mode=open|start|stop|toggle.

• Stop microphone capture and close the stream automatically after a live connection error.

• Start V1 streaming audio at sequence 2 because the full client request already consumes sequence 1.

• Bundle the Node WebSocket client required to attach Doubao authentication headers from Obsidian Desktop.

• Right-click supported audio files in the file explorer to transcribe them.

• Detect new Obsidian Recording … attachments when automatic transcription is enabled.

• Convert M4A, WebM, FLAC, AAC, and other browser-decodable audio to 16 kHz mono PCM WAV before calling the Doubao flash endpoint.

• Show live microphone transcription in a Crisp-styled right sidebar and append the final result to the note that was active when recording started.

• Store the API key through Obsidian SecretStorage rather than plugin data.

The plugin sends selected files and explicitly authorized microphone or computer audio directly to Volcengine. It does not run a proxy service.