feat: add wlk dictate push-to-talk command

README.md

@@ -386,6 +386,46 @@ async def websocket_endpoint(websocket: WebSocket):
 **Frontend Implementation**: The package includes an HTML/JavaScript implementation [here](https://github.com/QuentinFuxa/WhisperLiveKit/blob/main/whisperlivekit/web/live_transcription.html). You can also import it using `from whisperlivekit import get_inline_ui_html` & `page = get_inline_ui_html()`
 
 
+## Dictation (`wlk dictate`)
+
+Push-to-talk dictation for Linux (X11): press F9, speak, press F9 again, and the text is pasted where your cursor is. Audio goes straight into the streaming pipeline while you speak, so the text is ready within about half a second of stopping. No server or port is opened. F8 pastes the last text again.
+
+```bash
+sudo apt install xdotool x11-utils libnotify-bin   # paste, terminal detection, notifications
+pip install -e ".[dictate]"
+wlk dictate --model large-v3 --backend faster-whisper --backend-policy localagreement --lan en \
+  --min-chunk-size 1.0 --buffer_trimming_sec 8 --no-vac --log-level INFO
+```
+
+- `--hotkey` and `--repaste-hotkey` change the keys (default `f9` and `f8`). All other options are the same as `wlk serve`.
+- `--min-chunk-size 1.0 --buffer_trimming_sec 8 --no-vac` keep transcription ahead of your speech. With the defaults it re-transcribes too often and falls behind, so you wait after stopping.
+- Terminals get Ctrl+Shift+V instead of Ctrl+V. Your clipboard is restored after pasting.
+- Recordings with no speech, and filler Whisper produces from silence ("Thank you for watching."), are not pasted.
+- On RTX 50-series GPUs, `--compute-type int8_float16` needs a recent cuBLAS: `pip install nvidia-cublas-cu12==12.9.2.10` and set `LD_LIBRARY_PATH` to its `nvidia/cublas/lib` folder. With it, `large-v3` keeps up on a laptop GPU.
+
+Run it as a systemd user service:
+
+```ini
+# ~/.config/systemd/user/wlk-dictate.service
+[Unit]
+Description=WhisperLiveKit dictation
+After=graphical-session.target
+PartOf=graphical-session.target
+
+[Service]
+ExecStart=/path/to/venv/bin/wlk dictate --model large-v3 --backend faster-whisper --backend-policy localagreement --lan en --min-chunk-size 1.0 --buffer_trimming_sec 8 --no-vac --log-level INFO
+Restart=on-failure
+RestartSec=5
+
+[Install]
+WantedBy=graphical-session.target
+```
+
+```bash
+systemctl --user daemon-reload && systemctl --user enable --now wlk-dictate
+```
+
+
 ## Parameters & Configuration
 
 
@@ -399,6 +439,7 @@ async def websocket_endpoint(websocket: WebSocket):
 | `--diarization` | Enable speaker identification | `False` |
 | `--backend-policy` | Streaming strategy: `1`/`simulstreaming` uses AlignAtt SimulStreaming, `2`/`localagreement` uses the LocalAgreement policy | `simulstreaming` |
 | `--backend` | ASR backend selector. `auto` picks MLX on macOS (if installed), otherwise Faster-Whisper, otherwise vanilla Whisper. Options: `mlx-whisper`, `faster-whisper`, `whisper`, `openai-api` (LocalAgreement only), `funasr` (SenseVoiceSmall, LocalAgreement only), `voxtral-mlx` (Apple Silicon), `voxtral` (HuggingFace), `qwen3-vllm`, `qwen3-vllm-metal` (Apple Silicon), `qwen3-streaming` (HuggingFace, CUDA/MPS/CPU), `canary` (NeMo, CUDA/CPU, LocalAgreement only) | `auto` |
+| `--compute-type` | CTranslate2 compute type for faster-whisper models, e.g. `float16`, `int8_float16`, `float32`. `auto` lets CTranslate2 choose | `auto` |
 | `--pause-segmentation-seconds` | Create a stable transcript boundary when a VAD pause is longer than this many seconds. Use `0` to disable pause-based segmentation. | `5.0` |
 | `--asr-coalesce-min-s` | Defer an ASR call until this much new audio has accrued. Trades update cadence for fewer encoder passes; use `0` to disable. | `0` |
 | `--no-vac` | Disable Voice Activity Controller. NOT ADVISED | `False` |

docs/configuration.md

@@ -12,6 +12,7 @@ same settings as the CLI.
 | `--model` | Model size or name for the selected backend | `base` |
 | `--language` | Source language; coverage depends on the backend | `auto` |
 | `--backend-policy` | Whisper streaming policy: `simulstreaming` or `localagreement` | `simulstreaming` |
+| `--compute-type` | CTranslate2 compute type for faster-whisper models, e.g. `float16` or `int8_float16` | `auto` |
 | `--target-language` | Enable translation to this language | disabled |
 | `--translation-backend` | `nllb`, the [AlignAtt4LLM server](translation-alignatt.md), or [local MLX translation](translation-mlx.md) | `nllb` |
 | `--diarization` | Attribute words to speakers | disabled |

pyproject.toml

@@ -47,6 +47,7 @@ test = [
     "matplotlib>=3.8",
 ]
 translation = ["nllw"]
+dictate = ["pynput", "sounddevice", "pyperclip"]
 sentence_tokenizer = ["mosestokenizer", "wtpsplit"]
 funasr = ["funasr~=1.4.1"]
 mlx-whisper = [

tests/test_compute_type.py

@@ -0,0 +1,86 @@
+"""--compute-type reaches every faster-whisper model WhisperLiveKit creates."""
+
+import sys
+
+import faster_whisper
+
+from tests.test_backends import _base_simul_kwargs, _patch_simul_loaders
+
+
+def test_cli_accepts_compute_type(monkeypatch):
+    from whisperlivekit.parse_args import parse_args
+
+    monkeypatch.setattr(sys, "argv", ["whisperlivekit-server", "--compute-type", "float16"])
+
+    assert parse_args().compute_type == "float16"
+
+
+def test_cli_compute_type_defaults_to_auto(monkeypatch):
+    from whisperlivekit.parse_args import parse_args
+
+    monkeypatch.setattr(sys, "argv", ["whisperlivekit-server"])
+
+    assert parse_args().compute_type == "auto"
+
+
+def test_simulstreaming_encoder_uses_compute_type(monkeypatch):
+    backend, _ = _patch_simul_loaders(monkeypatch)
+
+    asr = backend.SimulStreamingASR(**_base_simul_kwargs(compute_type="float16"))
+
+    assert asr.fw_encoder.kwargs["compute_type"] == "float16"
+
+
+def test_simulstreaming_encoder_compute_type_defaults_to_auto(monkeypatch):
+    backend, _ = _patch_simul_loaders(monkeypatch)
+
+    asr = backend.SimulStreamingASR(**_base_simul_kwargs())
+
+    assert asr.fw_encoder.kwargs["compute_type"] == "auto"
+
+
+def test_faster_whisper_asr_uses_compute_type(monkeypatch):
+    from whisperlivekit.local_agreement.backends import FasterWhisperASR
+
+    created = {}
+
+    class FakeWhisperModel:
+        def __init__(self, model_size_or_path, **kwargs):
+            created.update(kwargs)
+
+    monkeypatch.setattr(faster_whisper, "WhisperModel", FakeWhisperModel)
+
+    FasterWhisperASR(lan="en", model_size="tiny", compute_type="float16")
+
+    assert created["compute_type"] == "float16"
+
+
+def test_backend_factory_passes_compute_type_to_faster_whisper(monkeypatch):
+    import whisperlivekit.local_agreement.whisper_online as whisper_online
+
+    created = {}
+
+    class FakeFasterWhisperASR:
+        def __init__(self, **kwargs):
+            created.update(kwargs)
+            self.transcribe_kargs = {}
+
+    monkeypatch.setattr(whisper_online, "FasterWhisperASR", FakeFasterWhisperASR)
+
+    whisper_online.backend_factory(
+        backend="faster-whisper",
+        lan="en",
+        model_size="tiny",
+        model_cache_dir=None,
+        model_dir=None,
+        model_path=None,
+        lora_path=None,
+        direct_english_translation=False,
+        buffer_trimming="segment",
+        buffer_trimming_sec=15,
+        confidence_validation=False,
+        warmup_file="",  # "" skips warmup; None downloads a sample file
+        compute_type="float16",
+    )
+
+    assert created["compute_type"] == "float16"