s2t 0.1.9__py3-none-any.whl → 0.1.11__py3-none-any.whl

s2t/cli.py

@@ -49,6 +49,7 @@ from .types import TranscriptionResult
 from .utils import (
     convert_wav_to_mp3,
     copy_to_clipboard,
+    debug_log,
     make_session_dir,
     open_in_shell_editor,
 )
@@ -57,6 +58,7 @@ from .whisper_engine import WhisperEngine
 
 def run_session(opts: SessionOptions) -> int:
     session_dir = make_session_dir(opts.outdir)
+    debug_log(opts.verbose, "cli", f"Session started; directory: {session_dir}")
     profile_data: dict = {}
     requested = opts.recording_format.lower()
     effective = requested
@@ -64,6 +66,12 @@ def run_session(opts: SessionOptions) -> int:
         logging.warning("ffmpeg not found; falling back to FLAC recording instead of MP3.")
         effective = "flac"
     ext = ".flac" if effective == "flac" else ".wav"
+    if requested != effective:
+        debug_log(
+            opts.verbose,
+            "cli",
+            f"Recording format adjusted: requested={requested}, effective={effective}",
+        )
 
     engine = WhisperEngine(
         model_name=opts.model,
@@ -77,6 +85,8 @@ def run_session(opts: SessionOptions) -> int:
         profile=profile_data if opts.profile else {},
     )
     ex, fut = engine.preload()
+    if ex is not None:
+        debug_log(opts.verbose, "cli", f"Model preload submitted for '{opts.model}'")
 
     # Determine translation target languages from options
     target_langs: list[str] = []
@@ -98,8 +108,14 @@ def run_session(opts: SessionOptions) -> int:
             detected_lang_event=detected_lang_event,
             detected_lang_holder=detected_lang,
         )
+        debug_log(
+            opts.verbose,
+            "cli",
+            f"Translation targets requested: {', '.join(target_langs)}",
+        )
 
-    tx_q: queue.Queue[tuple[int, Path, int, float]] = queue.Queue()
+    # Include split cause per chunk: "space" (manual), "pause" (auto), "finish" (final)
+    tx_q: queue.Queue[tuple[int, Path, int, float, str]] = queue.Queue()
     cumulative_text = ""
     next_to_emit = 1
     pending: dict[int, str] = {}
@@ -148,17 +164,29 @@ def run_session(opts: SessionOptions) -> int:
 
     def tx_worker():
         model = engine.resolve_model(fut)
+        debug_log(opts.verbose, "cli", "Transcription worker started")
         nonlocal cumulative_text, next_to_emit
         finished_texts: dict[int, str] = {}
+        causes: dict[int, str] = {}
         while True:
-            idx, path, frames, offset = tx_q.get()
+            idx, path, frames, offset, cause = tx_q.get()
             if idx == -1:
                 break
+            debug_log(
+                opts.verbose,
+                "cli",
+                f"Dequeued chunk {idx}: {path.name if path else '(final)'} (frames={frames}, offset={offset:.3f}, cause={cause or '-'})",
+            )
             # If in spoken-prompt mode, ensure we don't process payload chunks before prompt is done
             if opts.prompt and idx > 1 and not prompt_done.is_set():
+                debug_log(opts.verbose, "cli", f"Waiting for prompt before processing chunk {idx}")
                 prompt_done.wait()
             # Build latest-ready prompt based on already finished chunks
             prompt = _build_latest_ready_prompt(idx, finished_texts)
+            if prompt:
+                debug_log(
+                    opts.verbose, "cli", f"Built initial prompt for chunk {idx} (len={len(prompt)})"
+                )
             res = engine.transcribe_chunk(model, path, frames, initial_prompt=prompt)
             # Record detected language once (for translator preload if needed)
             if target_langs and detected_lang["code"] is None:
@@ -166,6 +194,7 @@ def run_session(opts: SessionOptions) -> int:
                 if lang_code:
                     detected_lang["code"] = lang_code
                     detected_lang_event.set()
+                    debug_log(opts.verbose, "cli", f"Detected source language: {lang_code}")
             engine.write_chunk_outputs(res, path)
             text_i = (res.get("text", "") or "").strip()
             with agg_lock:
@@ -174,20 +203,58 @@ def run_session(opts: SessionOptions) -> int:
                 results.append(res)
                 offsets.append(offset)
                 pending[idx] = text_i
+                # Track cause for formatting when emitting in-order
+                # cause is one of: "space", "pause", "finish" (or empty for sentinel)
+                # Default to "pause" if unknown to avoid extra blank lines.
+                causes[idx] = cause or "pause"
                 while next_to_emit in pending:
                     out = pending.pop(next_to_emit)
+                    cause_i = causes.get(next_to_emit) or "pause"
                     if out:
+                        # Live stdout behavior
                         print(out)
-                        print("")
-                        cumulative_text += out if not cumulative_text else ("\n\n" + out)
-                        try:
-                            copy_to_clipboard(cumulative_text)
-                        except Exception:
-                            pass
+                        if cause_i == "space":
+                            print("")  # blank line after SPACE
+                        # Build cumulative text with post-separator semantics
+                        if not cumulative_text:
+                            cumulative_text = out
+                        else:
+                            cumulative_text += out
+                        # Append separator AFTER the chunk, matching stdout
+                        if cause_i == "space":
+                            if not cumulative_text.endswith("\n\n"):
+                                # ensure exactly one paragraph break
+                                if cumulative_text.endswith("\n"):
+                                    cumulative_text += "\n"
+                                else:
+                                    cumulative_text += "\n\n"
+                        else:
+                            # single line break after non-space chunks
+                            if not (
+                                cumulative_text.endswith("\n") or cumulative_text.endswith("\n\n")
+                            ):
+                                cumulative_text += "\n"
+                    else:
+                        # Even if chunk text is empty, respect SPACE as a paragraph break
+                        if cause_i == "space":
+                            print("")  # blank line on stdout
+                            if cumulative_text:
+                                if cumulative_text.endswith("\n\n"):
+                                    pass
+                                elif cumulative_text.endswith("\n"):
+                                    cumulative_text += "\n"
+                                else:
+                                    cumulative_text += "\n\n"
+                        # For empty non-space chunks, do not alter cumulative_text
+                    try:
+                        copy_to_clipboard(cumulative_text)
+                    except Exception:
+                        pass
                     next_to_emit += 1
                 # If this was the prompt chunk, signal readiness and instruct user
                 if opts.prompt and idx == 1 and not prompt_done.is_set():
                     prompt_done.set()
+                    debug_log(opts.verbose, "cli", "Prompt transcribed; resuming payload")
                     print("=" * 60)
                     print("Prompt transcribed. Start speaking your main content now.")
                     print("=" * 60)
@@ -195,6 +262,7 @@ def run_session(opts: SessionOptions) -> int:
                     if prompt_resume_event is not None:
                         prompt_resume_event.set()
         tx_done.set()
+        debug_log(opts.verbose, "cli", "Transcription worker finished")
 
     tx_t = threading.Thread(target=tx_worker, daemon=True)
     tx_t.start()
@@ -202,6 +270,7 @@ def run_session(opts: SessionOptions) -> int:
     if opts.prompt:
         print("Prompt mode enabled: Speak your prompt first, then press SPACE.")
         print("Recording will wait for the prompt transcription before starting payload.")
+        debug_log(opts.verbose, "cli", "Prompt mode enabled")
     # Prepare resume event to pause recording between prompt and payload
     prompt_resume_event = threading.Event() if opts.prompt else None
     rec = Recorder(
@@ -221,6 +290,9 @@ def run_session(opts: SessionOptions) -> int:
     t1 = time.perf_counter()
     if opts.profile:
         profile_data["recording_sec"] = t1 - t0
+    debug_log(
+        opts.verbose, "cli", f"Recording finished in {(t1 - t0):.3f}s (chunks={len(chunk_paths)})"
+    )
     tx_t.join()
 
     merged: TranscriptionResult = engine.merge_results(results, chunk_offsets, cumulative_text)
@@ -230,8 +302,7 @@ def run_session(opts: SessionOptions) -> int:
     try:
         if chunk_paths:
             concat_audio(chunk_paths, base_audio_path, opts.rate, opts.channels)
-            if opts.verbose:
-                print(f"Merged audio written: {base_audio_path.name}", file=sys.stderr)
+            debug_log(opts.verbose, "cli", f"Merged audio written: {base_audio_path.name}")
             if requested == "mp3" and shutil.which("ffmpeg") is not None:
                 mp3_out = session_dir / "recording.mp3"
                 convert_wav_to_mp3(
@@ -242,11 +313,9 @@ def run_session(opts: SessionOptions) -> int:
                     ),
                     mp3_out,
                 )
-                if opts.verbose:
-                    print(f"Converted merged audio to MP3: {mp3_out.name}", file=sys.stderr)
+                debug_log(opts.verbose, "cli", f"Converted merged audio to MP3: {mp3_out.name}")
     except Exception as e:
-        if opts.verbose:
-            print(f"Warning: failed to merge chunk audio: {e}", file=sys.stderr)
+        debug_log(opts.verbose, "cli", f"Warning: failed to merge chunk audio: {e}")
 
     # Optionally delete chunk files (audio + per-chunk outputs)
     if chunk_paths and not opts.keep_chunks:
@@ -299,11 +368,11 @@ def run_session(opts: SessionOptions) -> int:
         # Decide source language: CLI hint takes precedence; else detected; else skip with warning
         src_lang = (opts.lang.lower() if opts.lang else (detected_lang["code"] or "")).strip()
         if not src_lang:
-            if opts.verbose:
-                print(
-                    "Warning: Could not determine source language for translation; skipping post-translation.",
-                    file=sys.stderr,
-                )
+            debug_log(
+                opts.verbose,
+                "cli",
+                "Warning: Could not determine source language for translation; skipping post-translation.",
+            )
         else:
             # Skip identical language targets
             effective_targets = [t for t in target_langs if t.lower() != src_lang.lower()]
@@ -335,8 +404,7 @@ def run_session(opts: SessionOptions) -> int:
                     f"{base_audio_path.stem}.{tgt}{base_audio_path.suffix}"
                 )
                 write_final_outputs(translated, session_dir, suffixed)
-                if opts.verbose:
-                    print(f"Created translated outputs for '{tgt}'.", file=sys.stderr)
+                debug_log(opts.verbose, "cli", f"Created translated outputs for '{tgt}'.")
             except Exception as e:
                 print(
                     f"Warning: failed to translate to '{tgt}': {e}",

s2t/recorder.py

@@ -11,6 +11,8 @@ from typing import Any, Protocol, cast, runtime_checkable
 
 import numpy as np
 
+from .utils import debug_log
+
 
 class Recorder:
     def __init__(
@@ -41,7 +43,7 @@ class Recorder:
 
     def run(
         self,
-        tx_queue: queue.Queue[tuple[int, Path, int, float]],
+        tx_queue: queue.Queue[tuple[int, Path, int, float, str]],
     ) -> tuple[list[Path], list[int], list[float]]:
         import platform
         import termios
@@ -71,14 +73,12 @@ class Recorder:
                     ms = cast(_MSVCRT, msvcrt)
 
                     last_space = 0.0
-                    if self.verbose:
-                        print("[key] using msvcrt (Windows)", file=sys.stderr)
+                    debug_log(self.verbose, "recorder", "Key input: using msvcrt (Windows)")
                     while not stop_evt.is_set():
                         if ms.kbhit():
                             ch = ms.getwch()
                             if ch in ("\r", "\n"):
-                                if self.verbose:
-                                    print("[key] ENTER", file=sys.stderr)
+                                debug_log(self.verbose, "recorder", "Key input: ENTER")
                                 evt_q.put("ENTER")
                                 break
                             if ch == " ":
@@ -88,8 +88,7 @@ class Recorder:
                                 ):
                                     continue
                                 last_space = now
-                                if self.verbose:
-                                    print("[key] SPACE", file=sys.stderr)
+                                debug_log(self.verbose, "recorder", "Key input: SPACE")
                                 evt_q.put("SPACE")
                         time.sleep(0.01)
                 else:
@@ -97,8 +96,9 @@ class Recorder:
                     try:
                         if sys.stdin.isatty():
                             fd = sys.stdin.fileno()
-                            if self.verbose:
-                                print("[key] using sys.stdin (isatty, fd read)", file=sys.stderr)
+                            debug_log(
+                                self.verbose, "recorder", "Key input: using sys.stdin (TTY fd read)"
+                            )
                             old = termios.tcgetattr(fd)
                             tty.setcbreak(fd)
                             last_space = 0.0
@@ -114,8 +114,7 @@ class Recorder:
                                             continue
                                         ch = ch_b.decode(errors="ignore")
                                         if ch in ("\n", "\r"):
-                                            if self.verbose:
-                                                print("[key] ENTER", file=sys.stderr)
+                                            debug_log(self.verbose, "recorder", "Key input: ENTER")
                                             evt_q.put("ENTER")
                                             break
                                         if ch == " ":
@@ -125,8 +124,7 @@ class Recorder:
                                             ):
                                                 continue
                                             last_space = now
-                                            if self.verbose:
-                                                print("[key] SPACE", file=sys.stderr)
+                                            debug_log(self.verbose, "recorder", "Key input: SPACE")
                                             evt_q.put("SPACE")
                             finally:
                                 termios.tcsetattr(fd, termios.TCSADRAIN, old)
@@ -137,8 +135,11 @@ class Recorder:
                             try:
                                 fd = os.open("/dev/tty", os.O_RDONLY)
                                 using_devtty = True
-                                if self.verbose:
-                                    print("[key] using /dev/tty (stdin not TTY)", file=sys.stderr)
+                                debug_log(
+                                    self.verbose,
+                                    "recorder",
+                                    "Key input: using /dev/tty (stdin not TTY)",
+                                )
                                 old = termios.tcgetattr(fd)
                                 tty.setcbreak(fd)
                                 last_space = 0.0
@@ -151,8 +152,9 @@ class Recorder:
                                                 continue
                                             ch = ch_b.decode(errors="ignore")
                                             if ch in ("\n", "\r"):
-                                                if self.verbose:
-                                                    print("[key] ENTER", file=sys.stderr)
+                                                debug_log(
+                                                    self.verbose, "recorder", "Key input: ENTER"
+                                                )
                                                 evt_q.put("ENTER")
                                                 break
                                             if ch == " ":
@@ -162,8 +164,9 @@ class Recorder:
                                                 ):
                                                     continue
                                                 last_space = now
-                                                if self.verbose:
-                                                    print("[key] SPACE", file=sys.stderr)
+                                                debug_log(
+                                                    self.verbose, "recorder", "Key input: SPACE"
+                                                )
                                                 evt_q.put("SPACE")
                                 finally:
                                     termios.tcsetattr(fd, termios.TCSADRAIN, old)
@@ -185,14 +188,16 @@ class Recorder:
                                         continue
                                     # If user hits Enter on empty line, treat as ENTER
                                     if line == "\n" or line == "\r\n":
-                                        if self.verbose:
-                                            print("[key] ENTER (line mode)", file=sys.stderr)
+                                        debug_log(
+                                            self.verbose, "recorder", "Key input: ENTER (line mode)"
+                                        )
                                         evt_q.put("ENTER")
                                         break
                                     # If first non-empty char is space, treat as SPACE
                                     if line and line[0] == " ":
-                                        if self.verbose:
-                                            print("[key] SPACE (line mode)", file=sys.stderr)
+                                        debug_log(
+                                            self.verbose, "recorder", "Key input: SPACE (line mode)"
+                                        )
                                         evt_q.put("SPACE")
                     except Exception as e:
                         print(f"Warning: key reader failed: {e}", file=sys.stderr)
@@ -224,7 +229,7 @@ class Recorder:
             threshold_rms = 0.015  # conservative RMS threshold for float32 [-1,1]
             split_cooldown_sec = 0.2
 
-            def _do_split() -> None:
+            def _do_split(cause: str) -> None:
                 nonlocal fh, frames_written, cur_path, chunk_index, offset_seconds_total
                 fh.flush()
                 fh.close()
@@ -234,12 +239,19 @@ class Recorder:
                     chunk_frames.append(frames_written)
                     chunk_offsets.append(offset_seconds_total)
                     offset_seconds_total += dur
-                    if self.verbose:
-                        print(
-                            f"Saved chunk: {cur_path.name} ({dur:.2f}s)",
-                            file=sys.stderr,
-                        )
-                    tx_queue.put((chunk_index, cur_path, frames_written, chunk_offsets[-1]))
+                    debug_log(
+                        self.verbose,
+                        "recorder",
+                        f"Saved chunk {chunk_index}: {cur_path.name} ({dur:.2f}s)",
+                    )
+                    # Include split cause so downstream can format output accordingly
+                    # cause: "space" (manual split) or "pause" (auto-split)
+                    tx_queue.put((chunk_index, cur_path, frames_written, chunk_offsets[-1], cause))
+                    debug_log(
+                        self.verbose,
+                        "recorder",
+                        f"Enqueued chunk {chunk_index} for transcription (cause={cause})",
+                    )
                 else:
                     try:
                         cur_path.unlink(missing_ok=True)
@@ -253,8 +265,14 @@ class Recorder:
                     and self.resume_event is not None
                 ):
                     self._paused = True
+                    debug_log(
+                        self.verbose,
+                        "recorder",
+                        "Paused after first chunk; waiting for resume (prompt mode)",
+                    )
                     self.resume_event.wait()
                     self._paused = False
+                    debug_log(self.verbose, "recorder", "Resumed after prompt")
                 cur_path = self.session_dir / f"chunk_{chunk_index:04d}{self.ext}"
                 fh = sf.SoundFile(
                     str(cur_path),
@@ -270,8 +288,13 @@ class Recorder:
                 try:
                     while True:
                         cmd = ctrl_q.get_nowait()
-                        if cmd == "split":
-                            _do_split()
+                        if cmd == "split_manual":
+                            _do_split("space")
+                            # Reset silence tracking on manual split
+                            silent_frames_run = 0
+                            seen_non_silent = False
+                        elif cmd == "split_auto":
+                            _do_split("pause")
                             # Reset silence tracking on manual split
                             silent_frames_run = 0
                             seen_non_silent = False
@@ -284,20 +307,33 @@ class Recorder:
                                 chunk_frames.append(frames_written)
                                 chunk_offsets.append(offset_seconds_total)
                                 offset_seconds_total += dur
-                                if self.verbose:
-                                    print(
-                                        f"Saved chunk: {cur_path.name} ({dur:.2f}s)",
-                                        file=sys.stderr,
-                                    )
+                                debug_log(
+                                    self.verbose,
+                                    "recorder",
+                                    f"Saved chunk {chunk_index}: {cur_path.name} ({dur:.2f}s)",
+                                )
+                                # Final chunk – mark cause as "finish" so downstream can avoid extra blank spacing
                                 tx_queue.put(
-                                    (chunk_index, cur_path, frames_written, chunk_offsets[-1])
+                                    (
+                                        chunk_index,
+                                        cur_path,
+                                        frames_written,
+                                        chunk_offsets[-1],
+                                        "finish",
+                                    )
+                                )
+                                debug_log(
+                                    self.verbose,
+                                    "recorder",
+                                    f"Enqueued final chunk {chunk_index} for transcription",
                                 )
                             else:
                                 try:
                                     cur_path.unlink(missing_ok=True)
                                 except Exception:
                                     pass
-                            tx_queue.put((-1, Path(), 0, 0.0))
+                            tx_queue.put((-1, Path(), 0, 0.0, ""))
+                            debug_log(self.verbose, "recorder", "Signaled transcription finish")
                             return
                 except queue.Empty:
                     pass
@@ -342,18 +378,18 @@ class Recorder:
                         enough_length = frames_written >= int(self.samplerate * self.min_chunk_sec)
                         cooldown_ok = (time.perf_counter() - last_split_time) >= split_cooldown_sec
                         if enough_silence and enough_length and seen_non_silent and cooldown_ok:
-                            if self.verbose:
-                                print(
-                                    f"[auto] split (≥{self.silence_sec:.2f}s silence)",
-                                    file=sys.stderr,
-                                )
+                            debug_log(
+                                self.verbose,
+                                "recorder",
+                                f"Auto-split (≥{self.silence_sec:.2f}s silence)",
+                            )
                             last_split_time = time.perf_counter()
-                            # Queue a split for the next control phase
-                            ctrl_q.put("split")
+                            # Queue an auto split for the next control phase
+                            ctrl_q.put("split_auto")
                             # Reset silence tracking now to avoid cascaded triggers
                             silent_frames_run = 0
                             seen_non_silent = False
-            tx_queue.put((-1, Path(), 0, 0.0))
+            tx_queue.put((-1, Path(), 0, 0.0, ""))
 
         def cb(indata: Any, frames: int, time_info: Any, status: Any) -> None:
             if status:
@@ -375,6 +411,12 @@ class Recorder:
         print("—" * 60)
         print("")
 
+        debug_log(
+            self.verbose,
+            "recorder",
+            f"Recording started (rate={self.samplerate}, channels={self.channels}, ext={self.ext})",
+        )
+
         import sounddevice as sd
 
         with sd.InputStream(samplerate=self.samplerate, channels=self.channels, callback=cb):
@@ -384,9 +426,10 @@ class Recorder:
                 except queue.Empty:
                     continue
                 if evt == "SPACE":
-                    ctrl_q.put("split")
+                    ctrl_q.put("split_manual")
                 elif evt == "ENTER":
                     ctrl_q.put("finish")
                     break
         writer_t.join()
+        debug_log(self.verbose, "recorder", "Recording finished")
         return chunk_paths, chunk_frames, chunk_offsets

s2t/translator/argos_backend.py

@@ -8,6 +8,7 @@ from collections.abc import Iterable
 from pathlib import Path
 
 from ..types import SegmentDict, TranscriptionResult
+from ..utils import debug_log
 
 # Global install coordination to avoid duplicate downloads in parallel
 _install_lock = threading.Lock()
@@ -25,8 +26,7 @@ class ArgosTranslator:
         self.verbose = verbose
 
     def _debug(self, msg: str) -> None:
-        if self.verbose:
-            print(msg)
+        debug_log(self.verbose, "argos", msg)
 
     @staticmethod
     def _guess_packages_dir() -> str:

s2t/utils.py

@@ -5,6 +5,7 @@ import platform
 import shutil
 import subprocess
 import sys
+import time
 from datetime import datetime
 from pathlib import Path
 
@@ -36,6 +37,25 @@ def convert_wav_to_mp3(wav_path: Path, mp3_path: Path) -> None:
     subprocess.run(cmd, check=True, stdout=subprocess.DEVNULL, stderr=subprocess.STDOUT)
 
 
+# Baseline at program start for relative timestamps in verbose logs
+_START_TIME = time.perf_counter()
+
+
+def debug_log(verbose: bool, component: str, msg: str) -> None:
+    """Emit a timestamped debug line to stderr if verbose is enabled.
+
+    Args:
+        verbose: Whether verbose mode is active.
+        component: Short component tag (e.g., 'recorder', 'whisper', 'cli', 'argos').
+        msg: Message to print.
+    """
+    if not verbose:
+        return
+    elapsed = time.perf_counter() - _START_TIME
+    # Elapsed with milliseconds precision
+    print(f"[+{elapsed:.3f}s] [{component}] {msg}", file=sys.stderr, flush=True)
+
+
 def copy_to_clipboard(text: str) -> None:
     system = platform.system()
     try:

s2t/whisper_engine.py

@@ -5,7 +5,20 @@ from concurrent.futures import Future, ThreadPoolExecutor
 from pathlib import Path
 from typing import Any
 
+import numpy as np
+
 from .types import SegmentDict, TranscriptionResult
+from .utils import debug_log
+
+# --- Tuning parameters (easy to adjust later) ---
+# Silence trim parameters operate on 16 kHz mono arrays
+TRIM_RMS_THRESHOLD: float = 0.012  # RMS threshold for speech vs. silence
+TRIM_MIN_VOICED_SEC: float = 0.5  # Require at least this much voiced audio to transcribe
+TRIM_PAD_MS: int = 50  # Keep a short pad around detected speech (ms)
+
+# Whisper inference behavior on low/empty audio
+WHISPER_NO_SPEECH_THRESHOLD: float = 0.7
+WHISPER_CONDITION_ON_PREV: bool = False
 
 
 class WhisperEngine:
@@ -48,6 +61,7 @@ class WhisperEngine:
                 return m, (t1 - t0)
 
             fut = self._executor.submit(_load, self.model_name)
+            debug_log(self.verbose, "whisper", f"Submitted model preload: {self.model_name}")
             return self._executor, fut
         except Exception:
             return None, None
@@ -62,6 +76,9 @@ class WhisperEngine:
                 self.profile["model_load_sec"] = self.profile.get("model_load_sec", 0.0) + float(
                     load_dur
                 )
+                debug_log(
+                    self.verbose, "whisper", f"Model resolved via preload in {float(load_dur):.3f}s"
+                )
             except Exception:
                 model = None
         if model is None:
@@ -69,6 +86,7 @@ class WhisperEngine:
             model = whisper.load_model(self.model_name)
             t1m = time.perf_counter()
             self.profile["model_load_sec"] = self.profile.get("model_load_sec", 0.0) + (t1m - t0m)
+            debug_log(self.verbose, "whisper", f"Loaded model synchronously in {(t1m - t0m):.3f}s")
         return model
 
     def transcribe_chunk(
@@ -79,10 +97,8 @@ class WhisperEngine:
         initial_prompt: str | None = None,
     ) -> TranscriptionResult:
         # Load audio without ffmpeg by reading via soundfile and passing a numpy array
-        # to Whisper. We ensure mono float32 at 16 kHz as expected by Whisper's API.
+        # to Whisper. Convert to mono float32 and resample to 16 kHz as expected by Whisper's API.
         task = "translate" if self.translate else "transcribe"
-        import numpy as np
-
         try:
             import soundfile as sf
         except Exception as e:
@@ -100,16 +116,71 @@ class WhisperEngine:
         # Resample to 16k expected by Whisper when passing arrays
         mono_16k: np.ndarray = resample_linear(mono, int(sr), 16000)
 
+        # Trim leading/trailing silence to avoid hallucinations on near-empty chunks
+        def _moving_rms(x: np.ndarray, win_len: int) -> np.ndarray:
+            if x.size == 0:
+                return np.zeros(0, dtype=np.float32)
+            win = np.ones(win_len, dtype=np.float32) / float(win_len)
+            sq = np.square(x.astype(np.float32, copy=False))
+            # same-length RMS via 'same' convolution
+            ma = np.convolve(sq, win, mode="same")
+            return np.sqrt(ma).astype(np.float32, copy=False)
+
+        def _trim_silence(x: np.ndarray, sr16k: int) -> tuple[np.ndarray, float, float]:
+            # Returns (trimmed, leading_sec, trailing_sec)
+            if x.size == 0:
+                return x, 0.0, 0.0
+            win_len = max(1, int(round(sr16k * 0.03)))  # 30 ms window
+            rms = _moving_rms(x, win_len)
+            thr = float(TRIM_RMS_THRESHOLD)
+            voiced = np.where(rms >= thr)[0]
+            if voiced.size == 0:
+                return np.zeros(0, dtype=np.float32), 0.0, float(x.size) / sr16k
+            start_idx = int(voiced[0])
+            end_idx = int(voiced[-1])
+            pad = int(round((TRIM_PAD_MS / 1000.0) * sr16k))
+            a = max(0, start_idx - pad)
+            b = min(x.size, end_idx + pad + 1)
+            lead_sec = float(a) / sr16k
+            trail_sec = float(x.size - b) / sr16k
+            return x[a:b], lead_sec, trail_sec
+
+        pre_sec = float(mono_16k.size) / 16000.0
+        trimmed, lead_sec, trail_sec = _trim_silence(mono_16k, 16000)
+        post_sec = float(trimmed.size) / 16000.0
+        debug_log(
+            self.verbose,
+            "whisper",
+            f"Chunk {audio_path.name}: trim {pre_sec:.2f}s -> {post_sec:.2f}s (lead {lead_sec:.2f}s, tail {trail_sec:.2f}s)",
+        )
+
+        # If too short after trimming, skip transcription
+        if post_sec < float(TRIM_MIN_VOICED_SEC):
+            debug_log(
+                self.verbose,
+                "whisper",
+                f"Chunk {audio_path.name}: too short after trim ({post_sec:.2f}s) – skipping",
+            )
+            return {"text": "", "segments": []}
+
         t0 = time.perf_counter()
+        debug_log(
+            self.verbose, "whisper", f"Transcribing chunk {audio_path.name} (frames={frames})"
+        )
         res: dict[str, Any] = model.transcribe(
-            mono_16k,
+            trimmed,
             task=task,
             language=self.language,
             fp16=False,
-            initial_prompt=initial_prompt,
+            initial_prompt=(initial_prompt if post_sec >= float(TRIM_MIN_VOICED_SEC) else None),
+            condition_on_previous_text=bool(WHISPER_CONDITION_ON_PREV),
+            no_speech_threshold=float(WHISPER_NO_SPEECH_THRESHOLD),
         )
         t1 = time.perf_counter()
         self.profile["transcribe_sec"] = self.profile.get("transcribe_sec", 0.0) + (t1 - t0)
+        debug_log(
+            self.verbose, "whisper", f"Transcribed chunk {audio_path.name} in {(t1 - t0):.3f}s"
+        )
         text_c = str(res.get("text", "") or "").strip()
         lang_code = str(res.get("language", "") or "")
         if self.native_segmentation:
@@ -117,8 +188,9 @@ class WhisperEngine:
             segs_typed: list[SegmentDict] = []
             for s in segs_raw:
                 try:
-                    start = float(s.get("start", 0.0))
-                    end = float(s.get("end", 0.0))
+                    # Adjust for leading trim so times align with original chunk timeline
+                    start = float(s.get("start", 0.0)) + float(lead_sec)
+                    end = float(s.get("end", 0.0)) + float(lead_sec)
                     text = str(s.get("text", "") or "")
                     segs_typed.append({"start": start, "end": end, "text": text})
                 except Exception:
@@ -129,8 +201,12 @@ class WhisperEngine:
             return out
         # Collapsed single segment per chunk
         segs_raw = res.get("segments", []) or []
-        start = float(segs_raw[0].get("start", 0.0)) if segs_raw else 0.0
-        end = float(segs_raw[-1].get("end", 0.0)) if segs_raw else (frames / float(self.samplerate))
+        start = (float(segs_raw[0].get("start", 0.0)) + float(lead_sec)) if segs_raw else 0.0
+        end = (
+            (float(segs_raw[-1].get("end", 0.0)) + float(lead_sec))
+            if segs_raw
+            else (frames / float(self.samplerate))
+        )
         out2: TranscriptionResult = {
             "text": text_c,
             "segments": ([{"start": start, "end": end, "text": text_c}] if text_c else []),
@@ -143,12 +219,17 @@ class WhisperEngine:
         try:
             from whisper.utils import get_writer
 
+            debug_log(self.verbose, "whisper", f"Writing outputs for {audio_path.name}")
             for fmt in ("txt", "srt", "vtt", "tsv", "json"):
                 writer = get_writer(fmt, str(self.session_dir))
                 writer(result, str(audio_path))
+            debug_log(self.verbose, "whisper", f"Wrote outputs for {audio_path.name}")
         except Exception as e:
-            if self.verbose:
-                print(f"Warning: failed to write chunk outputs for {audio_path.name}: {e}")
+            debug_log(
+                self.verbose,
+                "whisper",
+                f"Warning: failed to write chunk outputs for {audio_path.name}: {e}",
+            )
 
     def merge_results(
         self, results: list[TranscriptionResult], offsets: list[float], cumulative_text: str

{s2t-0.1.9.dist-info → s2t-0.1.11.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: s2t
-Version: 0.1.9
+Version: 0.1.11
 Summary: Speech to Text (s2t): Record audio, run Whisper, export formats, and copy transcript to clipboard.
 Author: Maintainers
 License-Expression: LicenseRef-Proprietary

s2t-0.1.11.dist-info/RECORD

@@ -0,0 +1,16 @@
+s2t/__init__.py,sha256=wV4E9i-7KrUn1dOtLUQB3ZGEKx9gRWH3hPHlpw-ZdWc,332
+s2t/cli.py,sha256=HgN-AXrBKn1MLjQd8U5hXBETlOKhOa8_3BH96GaVRss,24259
+s2t/config.py,sha256=uw4CZSSXmUvnlOrqBGR1Rcq-WdXucHj3KICRcCb_pkU,485
+s2t/outputs.py,sha256=Lo8VcARZ7QPuuQQNu8myD5J4c4NO1Rs0L1DLnzLe9tM,1546
+s2t/py.typed,sha256=AbpHGcgLb-kRsJGnwFEktk7uzpZOCcBY74-YBdrKVGs,1
+s2t/recorder.py,sha256=VYFqZ6LlP2zbwjWclZqM8ve5HEnZ3oyM9hLV1V3IkPI,20495
+s2t/types.py,sha256=jBiRN-tr0qVw-lhaXvnsyKrVGDyLkqEbxs9qkQ6qGqI,339
+s2t/utils.py,sha256=p7klapPW3Multxk261NlPtEpnEi3kpiTSHBPBTv4XC0,4059
+s2t/whisper_engine.py,sha256=T4HYmr1czwj78LsUdgRGWEBCfaghvHVqplQDaQDaR4o,10373
+s2t/translator/__init__.py,sha256=K-MKves7kZ4-62POfrmWeOcBaTjsTzeFSu8QNHqYuus,239
+s2t/translator/argos_backend.py,sha256=hXzQ8ZgJJOcUhcTJdTdVw1lSzptl8FXtfYr5PyOxKkg,19096
+s2t-0.1.11.dist-info/METADATA,sha256=kKEruaZB9lMsAeeejO4wwsj52jffj8tBCnp5WpYAlms,5475
+s2t-0.1.11.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+s2t-0.1.11.dist-info/entry_points.txt,sha256=JISIUlZAJ3DX1dB6zT3X_E3vcXI-eWEQKwHiT35fPKs,37
+s2t-0.1.11.dist-info/top_level.txt,sha256=o8N0JcuHdIrfX3iGHvntHiDC2XgN7__joyNu08ZOh0s,4
+s2t-0.1.11.dist-info/RECORD,,

s2t-0.1.9.dist-info/RECORD

@@ -1,16 +0,0 @@
-s2t/__init__.py,sha256=wV4E9i-7KrUn1dOtLUQB3ZGEKx9gRWH3hPHlpw-ZdWc,332
-s2t/cli.py,sha256=W-QyO8NhzslwK8cyodKyg9crXeffD1IlblcIGGqZt7Q,20572
-s2t/config.py,sha256=uw4CZSSXmUvnlOrqBGR1Rcq-WdXucHj3KICRcCb_pkU,485
-s2t/outputs.py,sha256=Lo8VcARZ7QPuuQQNu8myD5J4c4NO1Rs0L1DLnzLe9tM,1546
-s2t/py.typed,sha256=AbpHGcgLb-kRsJGnwFEktk7uzpZOCcBY74-YBdrKVGs,1
-s2t/recorder.py,sha256=tVUGwnHnlwQM9UtdHUlQIZ3GubCrop9suZYxc9yHHmE,18370
-s2t/types.py,sha256=jBiRN-tr0qVw-lhaXvnsyKrVGDyLkqEbxs9qkQ6qGqI,339
-s2t/utils.py,sha256=YU6YhiuONmqhrKte4DY5tiC5PP-yFExJMMBzFUiA8qA,3416
-s2t/whisper_engine.py,sha256=x-V7ST9e3JnwMWdbMh4C7dHjA420jaOtXH2-igeh7vc,6492
-s2t/translator/__init__.py,sha256=K-MKves7kZ4-62POfrmWeOcBaTjsTzeFSu8QNHqYuus,239
-s2t/translator/argos_backend.py,sha256=VW_OYFFBuNZgcWM-fbvR6XGokuxS2fptkCMFIO9MD1I,19068
-s2t-0.1.9.dist-info/METADATA,sha256=eY-j5C7WKtOhnxdKsFNEFU6MiT5G1cXJQpMroKAsIi8,5474
-s2t-0.1.9.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-s2t-0.1.9.dist-info/entry_points.txt,sha256=JISIUlZAJ3DX1dB6zT3X_E3vcXI-eWEQKwHiT35fPKs,37
-s2t-0.1.9.dist-info/top_level.txt,sha256=o8N0JcuHdIrfX3iGHvntHiDC2XgN7__joyNu08ZOh0s,4
-s2t-0.1.9.dist-info/RECORD,,