python-voiceio 0.6.0__py3-none-any.whl → 0.7.0__py3-none-any.whl

{python_voiceio-0.6.0.dist-info → python_voiceio-0.7.0.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: python-voiceio
-Version: 0.6.0
+Version: 0.7.0
 Summary: Voice dictation for Linux. Speak → text, locally, instantly.
 Author: Hugo Montenegro
 License-Expression: MIT

{python_voiceio-0.6.0.dist-info → python_voiceio-0.7.0.dist-info}/RECORD

@@ -1,5 +1,5 @@
-python_voiceio-0.6.0.dist-info/licenses/LICENSE,sha256=Gz61o8jFTAvZUZyB3nWDB3DQQVuipjfPkVu9W8hBHM0,1072
-voiceio/__init__.py,sha256=cID1jLnC_vj48GgMN6Yb1FA3JsQ95zNmCHmRYE8TFhY,22
+python_voiceio-0.7.0.dist-info/licenses/LICENSE,sha256=Gz61o8jFTAvZUZyB3nWDB3DQQVuipjfPkVu9W8hBHM0,1072
+voiceio/__init__.py,sha256=RaANGbRu5e-vehwXI1-Qe2ggPPfs1TQaZj072JdbLk4,22
 voiceio/__main__.py,sha256=xT5QCGGreYMisHO7Lh_Y-xAQ2TOzG2D7npKknGNcWSY,53
 voiceio/app.py,sha256=4o3nOv_Z_NCBr6KTHr1eSGT159ZE98dVBAy5wUOpzAo,64398
 voiceio/audit.py,sha256=vX3x-YAhmt6EZYRZd0_6wvTquZu5sxGMvH45BhSBnCk,14915
@@ -9,7 +9,7 @@ voiceio/backends.py,sha256=-rs1YhPblub4NjFb6qCD7Ih9FqBRXtPuvjVv9bQoQ1w,368
 voiceio/cli.py,sha256=_HfcWQt2Jf34YfI4W1RavsyqegQy9jrg0nwLe6DvD6E,71639
 voiceio/clipboard_read.py,sha256=CqbHSvOztcQ3u2I2EzFm0jdcDVKv56gl0T1YfWxNpmI,3390
 voiceio/commands.py,sha256=Vhtn8s5G5OcCbRN44YZ-_14Fa85ISXl_3Ix-KqSHs2A,4694
-voiceio/config.py,sha256=AOFaJras9rUPRDz7xeaZd4-WbWLOPUe0f0hYcQNwdJM,14042
+voiceio/config.py,sha256=CHDtEigtv-aKPsDFFGkc1AiqmubiFHeU1HaEaPgoIMc,14343
 voiceio/consent.py,sha256=jtdp1TkTbicUSTCDmV4qo4SdbS4kQpYzy1Kjc_7pJVw,1883
 voiceio/corrections.py,sha256=7moKhkVvo5CZU1mZobV2-MajkcHiwliKBY1dSBFh0w0,7445
 voiceio/demo.py,sha256=QRNJuNObXCsjwGD-Cz3dh81cfb21xiuSeVyyU720U9Q,8673
@@ -18,7 +18,7 @@ voiceio/health.py,sha256=5YNu1mMzwTo6I_Z_Mu3Bs46ZHDpVmxP0AXyMNAbaczc,15707
 voiceio/hints.py,sha256=dmoiMAJ0YGTWwFc9kIPvEs_BQJN7UEUiwYuLUD_GoQQ,1415
 voiceio/history.py,sha256=v7ciyZui8MndaKtTF5TdTKmql2ALOZNYKguE0Mo_lTY,5581
 voiceio/llm.py,sha256=_m13jodMVB1iEtze3VGkBCafkg85OssTPBEVyaRabgA,9250
-voiceio/llm_api.py,sha256=ex1brY10a8xNtw4cHCYC0yCth3Z02NkCQaHdB4XQPag,7497
+voiceio/llm_api.py,sha256=1wLLi6pliNmg-gS-bUKgwsGQSsLWAhX2XkEWXk292V8,9441
 voiceio/numbers.py,sha256=MP8jag4_F2OqUU5H14KzOmR9L8gWg48yQhLoUbGg_YI,8376
 voiceio/pidlock.py,sha256=RInlF_8H7wR0nqkdlhUfxYs7OGzHIvb5p-XDjjQSTNg,619
 voiceio/platform.py,sha256=f_8MmC5q1sCEuEPyKjRqlVvV9FExuGmwn_aJXhkpiEk,11809
@@ -29,7 +29,7 @@ voiceio/recorder.py,sha256=p1r0mH_plBD3BHqBbIGV_m2N4Mj5RAF7LfMu8u6mk5I,12813
 voiceio/retention.py,sha256=i_VWPCyykT4kQ1uKilTplpiIrACTYljn09BU1vL71A0,6019
 voiceio/service.py,sha256=jqX1opZyQh_dOJMK4bVf8Aoei0zOuuk1Ogta9BXTWGY,12106
 voiceio/snapshots.py,sha256=bm-lzIeYucDUDlPBZiFiWd-jyB2I6Ft1jRANZCO6go4,2351
-voiceio/streaming.py,sha256=XpF4fALTAgYqRB8zoDL81TuKlKLmUOc5_Cfd3auZFBI,21802
+voiceio/streaming.py,sha256=3JMXBU1DAIDvItGRKlh7RGlD1zcGwgNcrFCKLkWdmTI,23215
 voiceio/transcriber.py,sha256=2ev_hwdusve7ZZqht9A9wtjjT72rcr7SuhmSIW8aClk,10721
 voiceio/vad.py,sha256=72_ICk4jSuSPYvTHdHYtHzY8W5uaGEOz5FTdqkolJi8,5327
 voiceio/vocabulary.py,sha256=eGV7QZs3tN6Cw5YpyoZsVEk2M5fJvie1n2XFi9CD4hw,6321
@@ -71,8 +71,8 @@ voiceio/typers/pynput_type.py,sha256=DTtkT59M-EKOAlTNkvcZrRvf_Fq52OpI8S5t_8S-QAY
 voiceio/typers/wtype.py,sha256=d1wG-HZdYDQxXIysV_Xvk9HixDpaJgl3VEoKbP5vhIs,1786
 voiceio/typers/xdotool.py,sha256=dh1zhPqUT8ihJahJ7ZKm6PtfYY087UAzADx664DvQOM,1356
 voiceio/typers/ydotool.py,sha256=dt0W9ot__W8LQG2so8vZLGIw9leK2Q6g3i-Voo0LXE4,5327
-python_voiceio-0.6.0.dist-info/METADATA,sha256=HdRa-d2IgSiL7VZxA7lcXavG0dGkqcKGf6Lmzdf9xxk,15877
-python_voiceio-0.6.0.dist-info/WHEEL,sha256=K260EYznzXsJYBQGqmI8VTxEdiZYNvDZwW9cBh9-_MA,91
-python_voiceio-0.6.0.dist-info/entry_points.txt,sha256=U64fA65zxzyLoC8bgbn2ztQVWHLsc0o0H0qCk1J9DMc,218
-python_voiceio-0.6.0.dist-info/top_level.txt,sha256=piwtn309lD6uexQyXdZ-efAVBJF9y6Wfr48Z-8zkNhg,8
-python_voiceio-0.6.0.dist-info/RECORD,,
+python_voiceio-0.7.0.dist-info/METADATA,sha256=hV7gmFjuAjAJpCb9O01qqobjZfk7qxk3Kjow7l5-oaI,15877
+python_voiceio-0.7.0.dist-info/WHEEL,sha256=K260EYznzXsJYBQGqmI8VTxEdiZYNvDZwW9cBh9-_MA,91
+python_voiceio-0.7.0.dist-info/entry_points.txt,sha256=U64fA65zxzyLoC8bgbn2ztQVWHLsc0o0H0qCk1J9DMc,218
+python_voiceio-0.7.0.dist-info/top_level.txt,sha256=piwtn309lD6uexQyXdZ-efAVBJF9y6Wfr48Z-8zkNhg,8
+python_voiceio-0.7.0.dist-info/RECORD,,

voiceio/__init__.py

@@ -1 +1 @@
-__version__ = "0.6.0"
+__version__ = "0.7.0"

voiceio/config.py

@@ -147,7 +147,12 @@ class PostCorrectConfig:
     model can be overridden here; empty falls back to the autocorrect model.
     """
     enabled: bool = False
-    model: str = ""              # empty = use [autocorrect].model
+    # Latency-critical (blocks the commit): default to a fast cheap model.
+    # Bake-off on the real correction prompt: gemini-2.5-flash-lite 0.44s
+    # with fixes identical to kimi-k2's 2.0s. OpenRouter model id — on a
+    # custom base_url set a model your endpoint serves (empty = use
+    # [autocorrect].model).
+    model: str = "google/gemini-2.5-flash-lite"
     timeout_secs: float = 8.0
     min_words: int = 4           # skip utterances shorter than this
 

voiceio/llm_api.py

@@ -5,16 +5,73 @@ local Ollama (via /v1/chat/completions), etc. Zero dependencies beyond stdlib.
 """
 from __future__ import annotations
 
+import http.client
+import io
 import json
 import logging
 import os
+import threading
 import urllib.error
-import urllib.request
+from urllib.parse import urlsplit
 
 from voiceio.config import AutocorrectConfig
 
 log = logging.getLogger(__name__)
 
+# Keep-alive connection pool: a fresh TLS handshake costs ~0.2-0.5s per call,
+# which lands directly on postcorrect's stop-to-commit latency. Connections
+# are checked OUT for the duration of a request and returned only on clean
+# completion, so a deadline-abandoned postcorrect thread can never share a
+# socket with a later call.
+_POOL_LOCK = threading.Lock()
+_CONN_POOL: dict[str, http.client.HTTPConnection] = {}
+
+
+def _post_json(url: str, headers: dict, body: dict, timeout: float) -> dict:
+    """POST JSON over a pooled keep-alive connection; return the parsed reply.
+
+    A stale pooled connection (server closed it while idle) gets one retry on
+    a fresh one. Non-2xx raises urllib.error.HTTPError so callers keep their
+    existing status-code handling.
+    """
+    parts = urlsplit(url)
+    pool_key = f"{parts.scheme}://{parts.netloc}"
+    path = parts.path or "/"
+    payload = json.dumps(body).encode()
+
+    for attempt in (1, 2):
+        with _POOL_LOCK:
+            conn = _CONN_POOL.pop(pool_key, None)
+        pooled = conn is not None
+        if conn is None:
+            cls = (http.client.HTTPSConnection if parts.scheme == "https"
+                   else http.client.HTTPConnection)
+            conn = cls(parts.hostname, parts.port, timeout=timeout)
+        elif conn.sock is not None:
+            conn.sock.settimeout(timeout)
+        try:
+            conn.request("POST", path, body=payload, headers=headers)
+            resp = conn.getresponse()
+            data = resp.read()
+        except (http.client.HTTPException, OSError):
+            conn.close()
+            if not pooled or attempt == 2:
+                raise
+            continue  # stale keep-alive — retry once on a fresh connection
+        if resp.status // 100 != 2:
+            conn.close()
+            raise urllib.error.HTTPError(
+                url, resp.status, resp.reason, dict(resp.getheaders()),
+                io.BytesIO(data),
+            )
+        with _POOL_LOCK:
+            if pool_key in _CONN_POOL:
+                conn.close()  # keep at most one idle connection per host
+            else:
+                _CONN_POOL[pool_key] = conn
+        return json.loads(data)
+    raise AssertionError("unreachable")
+
 
 _LOCAL_HOSTS = ("localhost", "127.0.0.1", "0.0.0.0", "[::1]", "::1")
 _consent_warned = False
@@ -94,11 +151,7 @@ def _anthropic_request(
         "anthropic-version": "2023-06-01",
     }
 
-    req = urllib.request.Request(
-        url, data=json.dumps(body).encode(), headers=headers, method="POST",
-    )
-    with urllib.request.urlopen(req, timeout=timeout) as resp:
-        data = json.loads(resp.read())
+    data = _post_json(url, headers, body, timeout)
     # Anthropic returns content as a list of blocks; some thinking models
     # may also include `thinking` blocks which we ignore.
     blocks = data.get("content") or []
@@ -137,11 +190,7 @@ def _openai_request(
         "Authorization": f"Bearer {api_key}",
     }
 
-    req = urllib.request.Request(
-        url, data=json.dumps(body).encode(), headers=headers, method="POST",
-    )
-    with urllib.request.urlopen(req, timeout=timeout) as resp:
-        data = json.loads(resp.read())
+    data = _post_json(url, headers, body, timeout)
     # Be defensive: thinking models (Kimi K2.6, GPT reasoning, etc.) can
     # return content=None when the answer is in `reasoning` / `reasoning_content`
     # instead. Also some malformed responses lack `choices` entirely.

voiceio/streaming.py

@@ -32,7 +32,10 @@ _FREEZE_SILENCE_WINDOW_SECS = 0.3
 # and the decode path normalizes audio anyway, so an absolute threshold
 # would read a quiet mic's speech as silence (mid-word cuts) and a hot
 # noise floor as speech (freeze never fires).
-_FREEZE_SILENCE_RATIO = 0.15
+# Calibrated on real recordings: between-words noise floor sits at
+# ~0.27-0.38 of overall tail RMS, speech windows at >=0.9. 0.6 separates
+# both with margin; 0.15 never fired on a mic with a normal noise floor.
+_FREEZE_SILENCE_RATIO = 0.6
 _FREEZE_SILENCE_FLOOR = 1e-4  # digital silence is always quiet
 # Whisper conditions on ~224 prompt tokens; more frozen context is wasted.
 _FREEZE_CONTEXT_CHARS = 400
@@ -158,6 +161,12 @@ class StreamingSession:
         self._frozen_samples = 0
         self._frozen_raw = ""
         self._frozen_segments: list[dict] = []
+        # Decode backpressure: when the CPU is contended, decodes slow down
+        # and 1s-tick interim passes would queue back-to-back, each covering
+        # a longer tail — a starvation spiral. Require at least as much NEW
+        # audio as the last decode took before starting another interim.
+        self._last_decode_end = 0     # samples covered by the last decode
+        self._last_decode_secs = 0.0  # how long that decode took
         # Raw (pre-pipeline) text + confidence of the final pass, for history
         self.raw_final_text: str | None = None
         self.final_latency: dict = {}
@@ -294,6 +303,9 @@ class StreamingSession:
         tail = audio[self._frozen_samples:]
         if not final and len(tail) < self._sample_rate * min_seconds:
             return  # not enough new audio since the last freeze
+        grown_secs = (len(audio) - self._last_decode_end) / self._sample_rate
+        if not final and grown_secs < max(min_seconds, self._last_decode_secs):
+            return  # backpressure: don't decode faster than we can keep up
 
         # Freeze when the tail has grown long AND ends at a speech pause
         # (never cut mid-word). The freeze pass gets beam search because its
@@ -333,6 +345,8 @@ class StreamingSession:
                 return
             tail_segments = list(getattr(self._transcriber, "last_segments", []))
         t_transcribe = time.monotonic() - t0
+        self._last_decode_end = len(audio)
+        self._last_decode_secs = t_transcribe
 
         self.trace.append({
             "t": round(time.monotonic() - self._t0, 2),
@@ -353,6 +367,13 @@ class StreamingSession:
             )
         raw = self._frozen_raw if freeze else _join_text(self._frozen_raw, tail_text)
 
+        # A slow interim decode can complete AFTER the user pressed stop
+        # (observed: 29s under CPU contention). Applying it would overwrite
+        # the preedit with stale partial text moments before the final
+        # commit rewrites it again — discard it. Freeze bookkeeping above
+        # is kept: it is beam-quality and shrinks the final tail.
+        if not final and self._stop_event.is_set():
+            return
         # An interim pass that decoded nothing must not touch the display:
         # rewriting to frozen-only text would visibly delete the un-frozen
         # words the previous pass just showed (transient decoder flake).