#304 - RESEARCH: local voice transcription and rewrite (Superwhisper-style) — on-device/on-server, no cloud (NO implementation) - kayg/calternal

kayg commented

2026-09-28 07:05:51 +00:00

Owner

Owner (2026-09-28): 'file another issue for local transcription / modification of voice notes into text notes (like superwhisper) but we need to run it locally! have a codex research the best way to do this whenever capacity is available.'

Context: calternal records voice notes in the composer and (soon) in the Calendar item preview (the attach/record issue). Recordings are attachments in Home. The owner wants them turned into text notes: a transcription, plus an optional 'modification' step (clean up filler words, punctuate, restructure into a note, apply a style), like Superwhisper's modes. It must run locally: on the calternal server (a Rust server in a rootless podman container; the production VM has 4 vCPU and 7.7 GiB RAM with no GPU) and/or on the user's device (browser WebGPU/WASM; a future native macOS app). Never a cloud API by default (DESIGN business model: local features are always included; model-backed features use the customer's own API key). Licence rule: AGPL-3.0-only project, so the dependencies must be AGPL-compatible and the model weights must allow commercial use (no non-commercial weights).

Research (web search, benchmarks, licences):

Speech-to-text: whisper.cpp (and whisper-rs), faster-whisper/CTranslate2, Distil-Whisper, Whisper large-v3-turbo, NVIDIA Parakeet/Canary (check the licences), Moonshine, Kyutai STT, Vosk, sherpa-onnx, transformers.js/WebGPU in the browser, Apple SpeechAnalyzer (macOS 26) for the future native app. For each: accuracy (WER on common benchmarks, English plus the owner's likely languages: English, Hindi and other Indian languages; code-switching), speed on a 4-vCPU CPU (real-time factor), RAM, model size, streaming support, timestamps, diarization, licence of code and weights.
'Modification' step: small local LLMs (Qwen, Llama, Gemma, Phi, Mistral: check the licences and commercial use) through llama.cpp/candle/ort; the CPU cost on the VM; the quality of cleanup; how Superwhisper, MacWhisper, Wispr Flow, Aqua, VoiceInk and Apple's features structure their 'modes'. Also the option to use the customer's own API key when they choose.
Where it runs: a server-side job queue (background, bounded concurrency, per-user quotas; must not hurt search latency, #258) vs in-browser vs the native app; a hybrid.
Product: auto-transcribe on save vs on demand; where the text goes (the log entry's linked note, a new note, inline under the recording); search indexing of transcripts (#258 targets); editing the transcript; privacy.
Output: docs/research/voice-transcription.md (ASD-STE100) with a comparison table, a recommendation, a sizing estimate for the production VM, and numbered grill questions (grilling format). Do NOT implement. Commit on the branch, push, and post a summary on this issue.

Owner (2026-09-28): 'file another issue for local transcription / modification of voice notes into text notes (like superwhisper) but we need to run it locally! have a codex research the best way to do this whenever capacity is available.' Context: calternal records voice notes in the composer and (soon) in the Calendar item preview (the attach/record issue). Recordings are attachments in Home. The owner wants them turned into text notes: a transcription, plus an optional 'modification' step (clean up filler words, punctuate, restructure into a note, apply a style), like Superwhisper's modes. It must run **locally**: on the calternal server (a Rust server in a rootless podman container; the production VM has 4 vCPU and 7.7 GiB RAM with no GPU) and/or on the user's device (browser WebGPU/WASM; a future native macOS app). Never a cloud API by default (DESIGN business model: local features are always included; model-backed features use the customer's own API key). Licence rule: AGPL-3.0-only project, so the dependencies must be AGPL-compatible and the model weights must allow commercial use (no non-commercial weights). Research (web search, benchmarks, licences): - Speech-to-text: whisper.cpp (and whisper-rs), faster-whisper/CTranslate2, Distil-Whisper, Whisper large-v3-turbo, NVIDIA Parakeet/Canary (check the licences), Moonshine, Kyutai STT, Vosk, sherpa-onnx, transformers.js/WebGPU in the browser, Apple SpeechAnalyzer (macOS 26) for the future native app. For each: accuracy (WER on common benchmarks, English plus the owner's likely languages: English, Hindi and other Indian languages; code-switching), speed on a 4-vCPU CPU (real-time factor), RAM, model size, streaming support, timestamps, diarization, licence of code and weights. - 'Modification' step: small local LLMs (Qwen, Llama, Gemma, Phi, Mistral: check the licences and commercial use) through llama.cpp/candle/ort; the CPU cost on the VM; the quality of cleanup; how Superwhisper, MacWhisper, Wispr Flow, Aqua, VoiceInk and Apple's features structure their 'modes'. Also the option to use the customer's own API key when they choose. - Where it runs: a server-side job queue (background, bounded concurrency, per-user quotas; must not hurt search latency, #258) vs in-browser vs the native app; a hybrid. - Product: auto-transcribe on save vs on demand; where the text goes (the log entry's linked note, a new note, inline under the recording); search indexing of transcripts (#258 targets); editing the transcript; privacy. Output: docs/research/voice-transcription.md (ASD-STE100) with a comparison table, a recommendation, a sizing estimate for the production VM, and numbered grill questions (grilling format). Do NOT implement. Commit on the branch, push, and post a summary on this issue.

kayg commented

2026-09-28 07:15:53 +00:00

Author

Owner

Starting research on job/voice-research from dev at c9ec6aff84. Scope is research only; I will not run Cargo or Bun builds. I will review the voice attachment flow, search latency constraints, queue behavior, model and code licenses, current benchmark evidence, and local runtime options before writing the recommendation.

Starting research on job/voice-research from dev at c9ec6aff84a9ec810124a2cfbdd07c189eea1b9c. Scope is research only; I will not run Cargo or Bun builds. I will review the voice attachment flow, search latency constraints, queue behavior, model and code licenses, current benchmark evidence, and local runtime options before writing the recommendation.

kayg commented

2026-09-28 07:30:45 +00:00

Author

Owner

Finding: the composer already records voice memos in the browser and saves them under Documents/Voice memos/. The recorder accepts WebM/Opus, MP4/AAC, or Ogg/Opus and caps a recording at 30 minutes (apps/web/src/lib/composer/recorder.svelte.ts; crates/calternal-notes-core/src/attachments.rs). I found no ASR route or audio-transcript indexing path. The existing SQLite Worker supports a separate per-kind concurrency cap and retry/dedup behavior, but the queue has no per-User CPU quota. Search is a hard constraint: the 2026-09-28 100k profile on the shared 8-vCPU dev host reports warm hybrid p95 2,666.51 ms and marks the 150 ms target not met; it did not measure the 4-vCPU production VM (docs/perf/2026-09-28.md). I will recommend an explicit, bounded background job and require a production-VM profile before setting any throughput promise.

kayg commented

2026-09-28 07:47:02 +00:00

Author

Owner

Research is complete. The report is in docs/research/voice-transcription.md.

Recommendation: start with on-demand, local server transcription as a bounded background job. Pilot Qwen3-ASR-0.6B INT8 through sherpa-onnx; compare it with Vosk Hindi large and Whisper large-v3-turbo on the same consented samples. Save the raw result separately from cleanup. Let the User explicitly create a Note linked to the voice memo by stable identity.

Evidence and limits: the current recorder caps voice memos at 30 minutes. The Worker concurrency cap is per process and does not give per-User CPU quotas. The current Search profile misses its hybrid p95 target on the shared 8-vCPU host; it did not measure the 4-vCPU production VM. No controlled RTF or peak-RSS result was found for candidate models on that VM. The report includes language coverage, license and model caveats, product patterns, sizing estimates, a profiling plan, and 11 owner questions.

Decisions recommended for owner confirmation: required languages and Hindi script; on-demand versus opt-in automatic transcription; transcript persistence as Derived data or an explicit Note; live text and diarization scope; cleanup modes and cloud opt-in; retention and per-User resource limits. No product design was treated as settled where the design doc is silent.

Branch: job/voice-research. Base SHA: c9ec6aff84a9ec810124a2cfbdd07c189eea1b9c. Head SHA: 6924fba9c879a3d421b45dd41a54b1c659eb8216. The branch was pushed to origin.

Gate output:

git diff --cached --check: no output; exit 0.
Cargo/Bun build and test gates: Not run (research-only issue scope; builds prohibited; no code changed).
cargo clean: Removed 4 files, 3.3KiB total.

Known gap: no model or production-VM benchmark was run. This issue expressly prohibits builds, and the host is loaded.

Research is complete. The report is in `docs/research/voice-transcription.md`. Recommendation: start with on-demand, local server transcription as a bounded background job. Pilot Qwen3-ASR-0.6B INT8 through `sherpa-onnx`; compare it with Vosk Hindi large and Whisper large-v3-turbo on the same consented samples. Save the raw result separately from cleanup. Let the User explicitly create a Note linked to the voice memo by stable identity. Evidence and limits: the current recorder caps voice memos at 30 minutes. The Worker concurrency cap is per process and does not give per-User CPU quotas. The current Search profile misses its hybrid p95 target on the shared 8-vCPU host; it did not measure the 4-vCPU production VM. No controlled RTF or peak-RSS result was found for candidate models on that VM. The report includes language coverage, license and model caveats, product patterns, sizing estimates, a profiling plan, and 11 owner questions. Decisions recommended for owner confirmation: required languages and Hindi script; on-demand versus opt-in automatic transcription; transcript persistence as Derived data or an explicit Note; live text and diarization scope; cleanup modes and cloud opt-in; retention and per-User resource limits. No product design was treated as settled where the design doc is silent. Branch: `job/voice-research`. Base SHA: `c9ec6aff84a9ec810124a2cfbdd07c189eea1b9c`. Head SHA: `6924fba9c879a3d421b45dd41a54b1c659eb8216`. The branch was pushed to `origin`. Gate output: - `git diff --cached --check`: no output; exit 0. - Cargo/Bun build and test gates: `Not run (research-only issue scope; builds prohibited; no code changed).` - `cargo clean`: `Removed 4 files, 3.3KiB total`. Known gap: no model or production-VM benchmark was run. This issue expressly prohibits builds, and the host is loaded.

kayg referenced this issue from a commit

2026-09-28 07:48:43 +00:00

Merge job/voice-research (#304 research: local voice transcription)

kayg referenced this issue

2026-09-28 07:57:58 +00:00

RESEARCH: mail plugin (from mailternal) — decisions, architecture options, grill questions (NO implementation) #257

kayg commented

2026-09-28 08:02:22 +00:00

Author

Owner

Owner decisions (2026-09-28) on docs/research/voice-transcription.md:

V1 English only at launch. The owner asks: 'are you sure Qwen is the correct approach there? not Parakeet or S1-mini by Superwhisper?' The earlier pick (Qwen3-ASR-0.6B) was driven by Hindi, which is no longer needed. Follow-up research required (below).
V2 not needed (English only).
V3 B: automatic transcription after every voice note, on by default, as a low-priority background job. Plus a new Settings → Admin → Jobs page (like Immich) that shows what is running and lets an admin pause, resume or stop queues and jobs (filed separately).
V4 live text is not needed for now. The owner can increase the VM core count, so research how speed scales with cores (4 / 8 / 16 vCPU) and recommend a size.
V5 where the transcript goes depends on where the recording is added:
- to an event / log entry → a linked note (the sub-bullet rule: a nested child bullet under the entry, with the note title as the link text) that holds Title + the voice recording + the transcription;
- to a task → the task note;
- to a note → that note itself (the recording and its transcription go into it).
  All of them appear on the Calendar.
V6 local only by default. Owner: 'voice is one area which we can go local.' (A cloud key stays out of scope for voice.)

Follow-up research (English only): compare NVIDIA Parakeet TDT 0.6B v2 (English) and v3, Whisper large-v3-turbo (whisper.cpp), Distil-Whisper, Moonshine, Kyutai STT, Qwen3-ASR-0.6B, and Superwhisper's own models (the owner named 'S1-mini'; verify that it exists, what it is, and whether it can be self-hosted or licensed at all; do not guess). For each: English WER (Open ASR Leaderboard and other sources), real-time factor on CPU at 4, 8 and 16 vCPU (published numbers first; say clearly which numbers are measured and which are estimated), RAM, a Rust path (sherpa-onnx, ort, whisper-rs, candle), punctuation and casing out of the box, long-form (30 min) handling, and the licence of the weights for commercial use (CC-BY-4.0 attribution rules for Parakeet, etc.). Recommend one model plus a VM size. Also propose a small measurement protocol (fixed English clips, interleaved runs) that a later measurement job can run on a quiet host.

Owner decisions (2026-09-28) on docs/research/voice-transcription.md: - V1 **English only** at launch. The owner asks: 'are you sure Qwen is the correct approach there? not Parakeet or S1-mini by Superwhisper?' The earlier pick (Qwen3-ASR-0.6B) was driven by Hindi, which is no longer needed. **Follow-up research required** (below). - V2 not needed (English only). - V3 **B: automatic transcription after every voice note, on by default**, as a low-priority background job. Plus a new **Settings → Admin → Jobs** page (like Immich) that shows what is running and lets an admin pause, resume or stop queues and jobs (filed separately). - V4 live text is not needed for now. The owner **can increase the VM core count**, so research how speed scales with cores (4 / 8 / 16 vCPU) and recommend a size. - V5 **where the transcript goes depends on where the recording is added**: - to an **event / log entry** → a **linked note** (the sub-bullet rule: a nested child bullet under the entry, with the note title as the link text) that holds **Title + the voice recording + the transcription**; - to a **task** → the **task note**; - to a **note** → that **note itself** (the recording and its transcription go into it). All of them appear on the Calendar. - V6 **local only by default.** Owner: 'voice is one area which we can go local.' (A cloud key stays out of scope for voice.) **Follow-up research (English only):** compare NVIDIA **Parakeet TDT 0.6B v2** (English) and **v3**, **Whisper large-v3-turbo** (whisper.cpp), **Distil-Whisper**, **Moonshine**, **Kyutai STT**, Qwen3-ASR-0.6B, and Superwhisper's own models (the owner named '**S1-mini**'; verify that it exists, what it is, and whether it can be self-hosted or licensed at all; do not guess). For each: English WER (Open ASR Leaderboard and other sources), real-time factor on CPU at 4, 8 and 16 vCPU (published numbers first; say clearly which numbers are measured and which are estimated), RAM, a Rust path (sherpa-onnx, ort, whisper-rs, candle), punctuation and casing out of the box, long-form (30 min) handling, and the **licence of the weights for commercial use** (CC-BY-4.0 attribution rules for Parakeet, etc.). Recommend one model plus a VM size. Also propose a small measurement protocol (fixed English clips, interleaved runs) that a later measurement job can run on a quiet host.

kayg referenced this issue

2026-09-28 08:02:23 +00:00

ADMIN: Settings → Admin → Jobs page (Immich-style) — see, pause, resume and stop background work #315

kayg commented

2026-09-28 08:03:45 +00:00

Author

Owner

Starting the English-only follow-up research on job/voice-models, based on dev at c99fc31dd9. I will verify the Parakeet, Superwhisper S1-mini, and Qwen comparison from primary sources, update docs/research/voice-transcription.md, and run no builds.

Starting the English-only follow-up research on job/voice-models, based on dev at c99fc31dd9e0f542a02956a787010bbf1250384c. I will verify the Parakeet, Superwhisper S1-mini, and Qwen comparison from primary sources, update docs/research/voice-transcription.md, and run no builds.

kayg commented

2026-09-28 08:04:30 +00:00

Author

Owner

Owner link: https://huggingface.co/superwhisper/s1-mini. Orchestrator check of the model card (2026-09-28): S1-mini is not an ASR model. It is Superwhisper's post-ASR text normalizer: a 596M-parameter causal LM fine-tuned from Qwen3-0.6B, English only (v1). It removes fillers, resolves self-corrections, applies punctuation and casing, and converts spoken forms (numbers, dates, currency, emails) to written form, with style control lines (casual, semi-casual, semi-formal, formal). Formats: BF16 safetensors plus GGUF quantizations (Q4_K_M = 462 MB) for llama.cpp, Ollama and vLLM; it runs on CPU. The card reports 94.8% token accuracy on 7,519 held-out English cases (Q4_K_M, greedy). Licence: 'Apache 2.0 + naming clause'; the name 'S1-mini' by 'Superwhisper' must be kept with exact capitalization wherever it is used.
Implication: the pipeline is ASR (Parakeet/Whisper, to be chosen) → S1-mini cleanup, both local. S1-mini is the candidate for the 'Clean up' step that the first research left unpromised. The follow-up must: evaluate S1-mini as the cleanup stage (CPU cost per minute of audio at 4/8/16 vCPU via llama.cpp and a Rust binding; memory next to the ASR model; the naming-clause obligations in the UI and the NOTICE file); decide whether the raw transcript is kept next to the cleaned one (proposal: keep both, show the cleaned text, 'Show raw' toggle); and map the style control lines to a Settings → Voice choice.

Owner link: https://huggingface.co/superwhisper/s1-mini. Orchestrator check of the model card (2026-09-28): **S1-mini is not an ASR model.** It is Superwhisper's **post-ASR text normalizer**: a 596M-parameter causal LM fine-tuned from Qwen3-0.6B, English only (v1). It removes fillers, resolves self-corrections, applies punctuation and casing, and converts spoken forms (numbers, dates, currency, emails) to written form, with style control lines (casual, semi-casual, semi-formal, formal). Formats: BF16 safetensors plus GGUF quantizations (Q4_K_M = 462 MB) for llama.cpp, Ollama and vLLM; it runs on CPU. The card reports 94.8% token accuracy on 7,519 held-out English cases (Q4_K_M, greedy). Licence: **'Apache 2.0 + naming clause'**; the name 'S1-mini' by 'Superwhisper' must be kept with exact capitalization wherever it is used. Implication: the pipeline is **ASR (Parakeet/Whisper, to be chosen) → S1-mini cleanup**, both local. S1-mini is the candidate for the 'Clean up' step that the first research left unpromised. The follow-up must: evaluate S1-mini as the cleanup stage (CPU cost per minute of audio at 4/8/16 vCPU via llama.cpp and a Rust binding; memory next to the ASR model; the naming-clause obligations in the UI and the NOTICE file); decide whether the raw transcript is kept next to the cleaned one (proposal: keep both, show the cleaned text, 'Show raw' toggle); and map the style control lines to a Settings → Voice choice.

kayg commented

2026-09-28 08:34:07 +00:00

Author

Owner

Follow-up research complete

Updated docs/research/voice-transcription.md with the English-only pipeline, model comparison, 4/8/16-vCPU guidance, measurement protocol, and owner decisions. No code changes or builds were made.

Recommendation

Use NVIDIA Parakeet TDT 0.6B v2 as the first ASR candidate. Its model card reports 6.05 mean WER on the Open ASR Leaderboard, built-in punctuation/capitalization/word times, and an official INT8 ONNX path with a Rust example through sherpa-onnx.
Run S1-mini by Superwhisper as a separate local text-cleanup stage after ASR. Its 94.8% token accuracy is from 7,519 held-out cleanup cases and is not ASR WER or calternal quality evidence. Keep raw and cleaned text. The research recommends cleanup on by default with a Settings → Voice opt-out, cleaned text shown first, and a Show raw control.
Use 4 vCPU / 7.7 GiB for a one-job pilot. Start production at 8 vCPU / 16 GiB with one low-priority active job. Use 16 vCPU / 32 GiB only if an 8-vCPU measurement shows a queue or latency need.

There is no published controlled 4/8/16-vCPU result for Parakeet v2 or S1-mini. The available CPU data is a third-party Parakeet v3 INT8 profile using another runtime on an 8-vCPU host: 20.3× RTF at 4 threads and 26.3× at 8 threads on a 30-second clip, with about 1.5 GB peak RSS. It is not a calternal VM result and does not establish 30-minute or linear scaling performance. The note records this limitation and a same-host measurement protocol.

Owner decisions recorded

English only; automatic transcription after a memo is saved; no live transcription; local-only processing; transcript destination follows the Event/Log entry, Task, or Note; 4/8/16-vCPU comparison; Admin Jobs UI is a separate issue.

The research recommendations still needing owner confirmation are 8 vCPU / 16 GiB as the initial production size, S1-mini cleanup by default, its style defaults, five-minute ASR chunks, sequential model loading, and a license review before redistributing S1-mini. Its naming and NOTICE requirements are called out in the note.

Commit and gates

Branch: job/voice-models
Head: 48a96f3e410ac5f567d8bbadec23d0ca21fd7afa (pushed)

Gate output:

$ git diff --check

(exit code 0; no stdout or stderr)

Cargo and web build/test gates were not run because this issue is research-only and prohibits builds. cargo clean output:

Removed 2 files, 636B total

Local dev was already at the branch base; git merge dev output was Already up to date.

## Follow-up research complete Updated [docs/research/voice-transcription.md](https://git.kayg.org/kayg/calternal/src/branch/job/voice-models/docs/research/voice-transcription.md) with the English-only pipeline, model comparison, 4/8/16-vCPU guidance, measurement protocol, and owner decisions. No code changes or builds were made. ### Recommendation - Use NVIDIA Parakeet TDT 0.6B v2 as the first ASR candidate. Its model card reports 6.05 mean WER on the Open ASR Leaderboard, built-in punctuation/capitalization/word times, and an official INT8 ONNX path with a Rust example through `sherpa-onnx`. - Run S1-mini by Superwhisper as a separate local text-cleanup stage after ASR. Its 94.8% token accuracy is from 7,519 held-out cleanup cases and is not ASR WER or calternal quality evidence. Keep raw and cleaned text. The research recommends cleanup on by default with a Settings → Voice opt-out, cleaned text shown first, and a Show raw control. - Use 4 vCPU / 7.7 GiB for a one-job pilot. Start production at 8 vCPU / 16 GiB with one low-priority active job. Use 16 vCPU / 32 GiB only if an 8-vCPU measurement shows a queue or latency need. There is no published controlled 4/8/16-vCPU result for Parakeet v2 or S1-mini. The available CPU data is a third-party Parakeet v3 INT8 profile using another runtime on an 8-vCPU host: 20.3× RTF at 4 threads and 26.3× at 8 threads on a 30-second clip, with about 1.5 GB peak RSS. It is not a calternal VM result and does not establish 30-minute or linear scaling performance. The note records this limitation and a same-host measurement protocol. ### Owner decisions recorded English only; automatic transcription after a memo is saved; no live transcription; local-only processing; transcript destination follows the Event/Log entry, Task, or Note; 4/8/16-vCPU comparison; Admin Jobs UI is a separate issue. The research recommendations still needing owner confirmation are 8 vCPU / 16 GiB as the initial production size, S1-mini cleanup by default, its style defaults, five-minute ASR chunks, sequential model loading, and a license review before redistributing S1-mini. Its naming and NOTICE requirements are called out in the note. ### Commit and gates Branch: `job/voice-models` Head: `48a96f3e410ac5f567d8bbadec23d0ca21fd7afa` (pushed) Gate output: ```text $ git diff --check (exit code 0; no stdout or stderr) ``` Cargo and web build/test gates were not run because this issue is research-only and prohibits builds. `cargo clean` output: ```text Removed 2 files, 636B total ``` Local `dev` was already at the branch base; `git merge dev` output was `Already up to date.`

kayg referenced this issue from a commit

2026-09-28 08:38:21 +00:00

Merge job/voice-models (#304 follow-up: English ASR + S1-mini cleanup)

kayg commented

2026-09-28 08:42:20 +00:00

Author

Owner

Owner decisions (2026-09-28) on the follow-up research (merged 82843b08):

Pipeline accepted: Parakeet TDT 0.6B v2 (sherpa-onnx INT8, CC-BY-4.0 attribution) → S1-mini by Superwhisper cleanup on the text (cleaned text by default, a 'Show raw' toggle; Settings → Voice: style / structure / context, defaults semi-formal / prose / general).
V7: accept S1-mini's licence (Apache-2.0 + naming clause). Name it exactly 'S1-mini by Superwhisper' in Settings → Voice and in the credits/NOTICE. Do not bundle the weights in the AGPL release.
Model downloads: on by default, never blocking. 'it should download by default and not block deployment or server start or ui interaction.' The server starts and serves normally; a background job (visible in Settings → Admin → Jobs, #315) downloads both models (pinned revision + SHA-256, like calternal-embed's manifest) after start-up, resumable, with retry/backoff. Until they are ready, recordings save as normal and their transcription waits in the queue with the state 'Waiting for the voice model (downloading 42%)'. Offline or blocked networks: transcription stays pending with a clear message; nothing else is affected.
Recording UI: the owner points to https://sveltebits.xyz/micro/voice-pill as the recording animation to reuse and refine with calternal's aesthetics (cherry-pick per CLAUDE.md, motion tokens, reduced motion, the #291 spring language). It applies to the shared recorder bar (composer + #303 item preview).
VM: 8 vCPU / 16 GiB is the recommended size for auto-transcription; a measurement job confirms it before any promise.

Owner decisions (2026-09-28) on the follow-up research (merged 82843b08): - **Pipeline accepted:** Parakeet TDT 0.6B v2 (sherpa-onnx INT8, CC-BY-4.0 attribution) → **S1-mini by Superwhisper** cleanup on the text (cleaned text by default, a 'Show raw' toggle; Settings → Voice: style / structure / context, defaults semi-formal / prose / general). - V7: **accept S1-mini's licence** (Apache-2.0 + naming clause). Name it exactly 'S1-mini by Superwhisper' in Settings → Voice and in the credits/NOTICE. **Do not bundle the weights** in the AGPL release. - **Model downloads: on by default, never blocking.** 'it should download by default and not block deployment or server start or ui interaction.' The server starts and serves normally; a background job (visible in Settings → Admin → Jobs, #315) downloads both models (pinned revision + SHA-256, like calternal-embed's manifest) after start-up, resumable, with retry/backoff. Until they are ready, recordings save as normal and their transcription waits in the queue with the state 'Waiting for the voice model (downloading 42%)'. Offline or blocked networks: transcription stays pending with a clear message; nothing else is affected. - **Recording UI:** the owner points to https://sveltebits.xyz/micro/voice-pill as the recording animation to reuse and refine with calternal's aesthetics (cherry-pick per CLAUDE.md, motion tokens, reduced motion, the #291 spring language). It applies to the shared recorder bar (composer + #303 item preview). - VM: 8 vCPU / 16 GiB is the recommended size for auto-transcription; a measurement job confirms it before any promise.

kayg commented

2026-09-28 08:44:10 +00:00

Author

Owner

Owner refinement (2026-09-28) on the recording UI: the mic icon expands inline into the animated recording pill. The mic button itself morphs (width morph with the #236/#291 spring, as the tab bar does) into the recording state in place: the waveform/level animation, the elapsed time, Stop/Keep/Discard. No overlay, no sheet, no separate bar, no extra chrome. Owner: 'no over-the-top crap. our design looks great as is.' Take only the voice-pill's core motion idea from sveltebits; keep calternal's existing pill look (the same PillGroup/glass, tokens, icon sizes). Collapse back into the mic on Stop/Keep/Discard. Reduced motion: an instant swap. The same component in the composer tools pill and the Calendar item-preview pill.

Owner refinement (2026-09-28) on the recording UI: **the mic icon expands inline into the animated recording pill**. The mic button itself morphs (width morph with the #236/#291 spring, as the tab bar does) into the recording state in place: the waveform/level animation, the elapsed time, Stop/Keep/Discard. No overlay, no sheet, no separate bar, no extra chrome. Owner: 'no over-the-top crap. our design looks great as is.' Take only the voice-pill's core motion idea from sveltebits; keep calternal's existing pill look (the same PillGroup/glass, tokens, icon sizes). Collapse back into the mic on Stop/Keep/Discard. Reduced motion: an instant swap. The same component in the composer tools pill and the Calendar item-preview pill.

kayg referenced this issue

2026-09-28 18:44:50 +00:00

VOICE: local transcription (Parakeet v2 → S1-mini by Superwhisper), background model download, inline mic morph #373

kayg referenced this issue

2026-09-28 18:46:50 +00:00

VOICE: local transcription (Parakeet v2 → S1-mini by Superwhisper), background model download, inline mic morph #373

kayg referenced this issue

2026-09-30 06:44:15 +00:00

VOICE: assistant, composer notes, dictation + global shortcuts, capture anywhere, local background transcription (design) #488

kayg referenced this issue

2026-09-30 07:02:03 +00:00

RESEARCH: Sign in with ChatGPT for the voice assistant (Realtime eligibility, client ID for self-hosters) #490