RECCoon 1.0.0: rebrand to Tom Cooks, Parakeet-only, GPLv3, F-Droid-ready

1 files changed, 113 insertions, 0 deletions

diff --git a/TODO.md b/TODO.md
index 1fa9671..76c4ea6 100644
--- a/TODO.md
+++ b/TODO.md
@@ -1,5 +1,118 @@
 # RECCoon TODO
 
+## Requested (2026-10-05)
+
+### 1. Compare transcription models (Whisper / Parakeet / Whistle)
+
+#### 1a. Parakeet TDT 0.6B v3 — **done, integrated**
+
+- [x] **Test NVIDIA Parakeet v3** (`csukuangfj/sherpa-onnx-nemo-parakeet-tdt-0.6b-v3-int8`).
+  - [x] Added `parakeet_baseline.py` to `tools/whistle-eval/` and wired it into
+        `evaluate.py`.
+  - [x] Result (same six clips): average WER **3.7%** vs Whisper base 11.4% and
+        tiny 13.9%. Italian **2.8%** (vs base 16.7%) and **0%** on the noisy
+        Italian clip. The only errors are number formatting.
+  - [x] sherpa-onnx already supports it (`model_type="nemo_transducer"`), so no new
+        runtime is needed.
+  - [x] Integrated into the app: `ModelRepository.Model.PARAKEET_V3`,
+        `TranscriptionEngine` builds an `OfflineTransducerModelConfig`, and the
+        player model spinner lists "Parakeet v3 (best)".
+  - [ ] Verify decoding and memory use on a real phone; the int8 model is
+        ~670 MB (~652 MB encoder) and needs a large-heap device.
+  - [x] Removed the Whisper tiny/base options: Parakeet v3 is now the only model
+        (and therefore the default). The language picker was removed too, since
+        Parakeet handles language by itself. The `Model` enum keeps room for more
+        models later.
+  - [ ] Consider a smaller multilingual Parakeet/Canary variant if device RAM is tight.
+
+#### 1b. Cactus Whistle — tested, not integrated
+
+- [~] **Check whether Cactus Whistle transcribes better than multilingual Whisper.**
+      Reference: <https://huggingface.co/mrfakename/whistle-ONNX> (ONNX export of
+      `Cactus-Compute/whistle`; supports en/de/fr/es/it/nl/pl). Motivation: Whisper
+      makes many mistakes, and a smaller model could replace the ~50 MB sherpa-onnx
+      AAR and fix the git push problem.
+  - [x] Build a desktop evaluation harness under `tools/whistle-eval/` (`setup.sh`,
+        `evaluate.py`, `whisper_baseline.py`, `build-data.mjs`).
+  - [x] Baseline + metric: the harness runs Whisper tiny/base through the same
+        sherpa-onnx 1.13.8 version the app bundles and reports WER.
+  - [x] First result (2026-10-05, six TTS clips, EN + IT): Whistle is **competitive
+        but not a clear win** — average WER 11.8% vs Whisper base 11.4% and tiny
+        13.9%. On Italian Whistle was best (16.1% vs base 16.7%, tiny 23.2%); on
+        English tiny was best. See `tools/whistle-eval/README.md`.
+  - [ ] Decide the Android port after running the harness on **real RECCoon
+        recordings** — the TTS result is too close to call.
+  - [x] Feasibility: Whistle is a custom *Needle* encoder/decoder, **not** Whisper, so
+        sherpa-onnx cannot load it. An Android port needs `onnxruntime-android` plus
+        the mel filterbank, BPE tokenizer and engram features from `js/` /
+        `whistle.pack`.
+  - [x] Distribution: `whistle.pack` is ~17 MB (fits GitHub); the fp32 ONNX data is
+        ~145 MB. `build-data.mjs` reconstructs the `.onnx.data` files from the pack
+        byte-exactly, so only the pack needs to ship.
+  - [x] Decision: **do not add Whistle to the model picker** — Parakeet v3 is both
+        more accurate and supported by sherpa-onnx, so the custom-architecture port
+        is not worth it.
+
+### 2. Compact (compressed) recordings
+
+- [~] **Switch on the main screen to save very small compressed files instead of WAV.**
+  - [x] Codec: AAC-LC in MP4 (`.m4a`) for universal `MediaCodec` support; bitrate is
+        48 kbps per channel (`AudioCompressor`).
+  - [x] Added a **Compressed (small M4A)** checkbox under the record button, persisted
+        as `compress_recordings`.
+  - [x] The temporary WAV is transcoded to `.m4a` after stopping and then deleted; the
+        M4A is saved to Downloads instead of the WAV.
+  - [x] Markers still work through the existing JSON sidecar (M4A has no `cue ` chunk).
+  - [x] `TranscriptionEngine` now streams through `PcmAudioSource`, which decodes
+        compressed files with `MediaExtractor`/`MediaCodec`, so transcription still
+        works on M4A.
+  - [ ] Measure the size win and the transcription-quality impact on a real device and
+        write the numbers here. Expected size: ~22 MB/hour mono instead of ~317 MB.
+
+### 3. Forced microphone input (BT-Mic-Force)
+
+- [~] **Integrate the BT-Mic-Force mic-forcing feature into RECCoon.**
+  - [x] Added a **Mic** spinner (Auto / Phone mic / detected inputs) to the main
+        screen.
+  - [x] The choice is routed into `WavRecorder` via `AudioRecord.setPreferredDevice`
+        and the `AudioManager` SCO mode (`setMode`, `startBluetoothSco`,
+        `setBluetoothScoOn`).
+  - [x] Added `MODIFY_AUDIO_SETTINGS`, `BLUETOOTH_CONNECT` and (pre-31)
+        `BLUETOOTH`/`BLUETOOTH_ADMIN` permissions; `BLUETOOTH_CONNECT` is requested
+        only when a Bluetooth input is selected.
+  - [x] Reused the existing `AudioRecord` capture path; no second foreground service.
+  - [ ] Verify that Bluetooth SCO actually routes audio on a real device.
+
+### 4. Live monitoring, output picker, and quick discard
+
+- [x] **Live monitoring** (`Monitor` checkbox + output picker). `LiveMonitor` plays the
+      captured PCM through `AudioTrack` on a dedicated thread with a bounded queue, so
+      a slow output never stalls the recording. Output devices are listed like the mic
+      picker (Auto / phone speaker / wired / USB / Bluetooth). Feedback warning when
+      monitoring is enabled.
+- [x] **Stop and discard.** A **Discard** button stops the recording and deletes the
+      temporary file without saving.
+- [ ] Verify latency and routing on real devices (wired, Bluetooth A2DP, BT LE).
+      Bluetooth A2DP adds unavoidable latency; note whether it is usable.
+- [ ] Consider requesting audio focus and muting monitoring automatically when the
+      selected output is the built-in speaker.
+
+### 5. Rotation feedback (Alberto, from 1.4.0-alpha)
+
+- [x] **Level meters look upside down when the phone is turned.** Cause: both
+      activities were locked to `screenOrientation="portrait"`, so nothing followed
+      the device.
+  - [x] Removed the portrait lock so the UI rotates with the device.
+  - [x] Added `configChanges="orientation|screenSize|..."` so a rotation does not
+        recreate the activity and interrupt an in-progress recording.
+  - [x] **L/R swapped in landscape.** Android exposes the two mic channels in a fixed
+        device order; it does not tell the app which physical mic is which, so this is
+        not reliably auto-detectable. Added a persisted **Swap L/R** checkbox that
+        exchanges the channels in `WavRecorder`, so both the meters and the saved WAV
+        match the user's left/right.
+  - [ ] Confirm on a real device that rotating mid-recording keeps the recording and
+        the meters upright, and that Swap L/R fixes the inverted meters.
+
 ## Build / distribution
 
 - [ ] **The sherpa-onnx binaries are too large to keep in git.** The Android