friend-assistant 0.1.40 → 0.1.41

package/dist/BUILD.json

@@ -1,5 +1,5 @@
 {
   "name": "friend-assistant",
-  "version": "0.1.40",
-  "builtAt": "2026-10-13T02:23:53.362Z"
+  "version": "0.1.41",
+  "builtAt": "2026-10-13T04:06:44.122Z"
 }

package/dist/friend.mjs

@@ -3337,7 +3337,7 @@ function sanitizeLook(raw) {
 var DEFAULT_OLLAMA = { url: "http://127.0.0.1:11434", model: "qwen2.5:7b", contextTokens: 8192, vision: false };
 var DEFAULT_VOICE = {
   enabled: true,
-  sttModel: "base",
+  sttModel: "small",
   ttsVoice: "en_US-lessac-medium",
   language: null,
   speakTyped: true,
@@ -11035,7 +11035,8 @@ function createGateway(deps) {
       const session = token ? auth.resolveSession(token) : null;
       if (!session) return reply.code(401).send({ error: "Unauthorized" });
       try {
-        return await deps.onVoiceUtterance(session, body, String(req.headers["content-type"] ?? "audio/webm"), req.ip);
+        const wake = req.query.wake === "1";
+        return await deps.onVoiceUtterance(session, body, String(req.headers["content-type"] ?? "audio/webm"), req.ip, wake);
       } catch (e) {
         if (e instanceof HttpError) return reply.code(e.status).send({ error: e.message });
         audit.log("owner", "voice.utterance_failed", { message: e.message });
@@ -13761,6 +13762,39 @@ var Enrollment = class {
   }
 };
 
+// ../server/src/voice/wake.ts
+var GREETINGS = /* @__PURE__ */ new Set(["hey", "hay", "hi", "hello", "hellow", "hallo", "hullo", "ok", "okay", "yo", "he", "\u0939\u0947", "\u0939\u093E\u092F", "\u0939\u0945\u0932\u094B", "\u0939\u0947\u0932\u094B", "\u0939\u0948\u0932\u094B", "\u0939\u0932\u094B", "\u0913\u0915\u0947"]);
+var WORD = /[\p{L}\p{M}\p{N}'’]+/gu;
+var GAP = /^[\s,.:;!?…\-–—]+/u;
+function distance(a, b) {
+  const row = Array.from({ length: b.length + 1 }, (_, i) => i);
+  for (let i = 1; i <= a.length; i++) {
+    let prev = row[0];
+    row[0] = i;
+    for (let j = 1; j <= b.length; j++) {
+      const keep = row[j];
+      row[j] = Math.min(row[j] + 1, row[j - 1] + 1, prev + (a[i - 1] === b[j - 1] ? 0 : 1));
+      prev = keep;
+    }
+  }
+  return row[b.length];
+}
+var sameWord = (heard, name) => heard === name || name.length >= 4 && distance(heard, name) <= 1;
+function matchWake(text3, name) {
+  const no = { matched: false, rest: "" };
+  const wanted = [...name.normalize("NFC").toLowerCase().matchAll(WORD)].map((m) => m[0]);
+  if (wanted.length === 0) return no;
+  const words = [...text3.normalize("NFC").matchAll(WORD)].map((m) => ({ low: m[0].toLowerCase(), end: m.index + m[0].length }));
+  let at = 0;
+  if (words[0] && GREETINGS.has(words[0].low)) at = 1;
+  for (let i = 0; i < wanted.length; i++) {
+    const heard = words[at + i];
+    if (!heard || !sameWord(heard.low, wanted[i])) return no;
+  }
+  const last = words[at + wanted.length - 1];
+  return { matched: true, rest: text3.normalize("NFC").slice(last.end).replace(GAP, "").trim() };
+}
+
 // ../server/src/voice/service.ts
 var VoiceError = class extends Error {
 };
@@ -14817,7 +14851,7 @@ async function createApp(opts) {
     audit,
     webDir: opts.webDir,
     voiceAvailable,
-    onVoiceUtterance: async (session, audio, mime, ip) => {
+    onVoiceUtterance: async (session, audio, mime, ip, wake = false) => {
       if (!voice2) throw new HttpError(503, "Voice is not available");
       if (agent.isBusy()) throw new HttpError(409, "Friend is busy with the last request. Try again in a moment.");
       const epoch = lockEpoch;
@@ -14835,7 +14869,13 @@ async function createApp(opts) {
         throw e;
       }
       if (epoch !== lockEpoch || !auth.sessionActive(session.id)) throw new HttpError(401, "Session ended");
-      if (!/[\p{L}\p{N}]/u.test(heard)) return { text: "" };
+      if (!/[\p{L}\p{N}]/u.test(heard)) return wake ? { text: "", wake: false } : { text: "" };
+      if (wake) {
+        const m = matchWake(heard, auth.assistantName());
+        if (!m.matched) return { text: "", wake: false };
+        heard = m.rest;
+        if (!/[\p{L}\p{N}]/u.test(heard)) return { text: "", wake: true };
+      }
       const crash = (e) => audit.log("system", "voice.turn_crash", { message: e.message });
       const who = enrolled ? judge(embedding) : { trust: "unverified" };
       const trust = who.trust;
@@ -14857,7 +14897,7 @@ async function createApp(opts) {
         broadcast({ type: "voice.heard", text: heard, trust: trust === "owner" ? "owner" : "unverified" });
         void runVoiceTurn(heard, ip, epoch, "owner", void 0, spoken).catch(crash);
       }
-      return { text: heard };
+      return wake ? { text: heard, wake: true } : { text: heard };
     },
     voiceEnroll: enrollment && voice2 ? {
       state: () => enrollment.state(),
@@ -15635,7 +15675,7 @@ var KEY_HELP = {
   allowedHosts: "extra host names that may reach Friend (for example my-pc), separated by commas",
   trustProxy: "true when Friend sits behind your own reverse proxy",
   "voice.enabled": "true or false: use speech at all",
-  "voice.sttModel": "listening model: tiny, base, small, \u2026 (bigger is more accurate and slower)",
+  "voice.sttModel": "listening model: tiny, base, small, large-v3-turbo \u2026 (bigger is more accurate and slower; Marathi and far-from-mic speech do best with small or large-v3-turbo)",
   "voice.ttsVoice": "speaking voice (easier: friend voice setup, or Settings)",
   "voice.language": "force a spoken language such as en, hi or mr, or empty to detect it (empty is best when you mix languages)",
   "voice.hindiVoice": "the voice that reads Hindi (Devanagari) text: hi_IN-priyamvada-medium, hi_IN-pratham-medium or hi_IN-rohan-medium; empty uses the main voice",
@@ -17804,7 +17844,7 @@ async function daemonCommand(args, io2, deps = realDeps) {
 import crossSpawn5 from "cross-spawn";
 var PACKAGE = "friend-assistant";
 var real = {
-  current: () => true ? "0.1.40" : "development",
+  current: () => true ? "0.1.41" : "development",
   latest: () => new Promise((resolve13) => {
     const child = crossSpawn5("npm", ["view", PACKAGE, "version", "--prefer-online"], { stdio: ["ignore", "pipe", "ignore"] });
     let out = "";
@@ -18092,7 +18132,7 @@ async function run(argv, io2) {
       case "version":
       case "--version":
       case "-v":
-        io2.out(true ? "0.1.40" : "development");
+        io2.out(true ? "0.1.41" : "development");
         return 0;
       case "lan":
         return await lanCommand(rest, io2);

package/dist/voice/friend_voice/__main__.py

@@ -18,7 +18,7 @@ def main() -> int:
     ap.add_argument("--port", type=int, default=0, help="0 picks a free port")
     ap.add_argument("--fake", action="store_true", help="use stub engines (tests)")
     ap.add_argument("--data-dir", default=None, help="folder holding downloaded models")
-    ap.add_argument("--stt-model", default="base")
+    ap.add_argument("--stt-model", default="small")
     ap.add_argument("--tts-voice", default="en_US-lessac-medium")
     args = ap.parse_args()
     token = args.token or os.environ.get("FRIEND_VOICE_TOKEN")

package/dist/voice/friend_voice/models.py

@@ -46,7 +46,7 @@ def ensure_speaker_model(directory: Path, url: str = SPEAKER_URL, sha256: str =
     return target
 
 
-def load_engines(data_dir: Optional[str], stt_model: str = "base", tts_voice: str = "en_US-lessac-medium") -> Engines:
+def load_engines(data_dir: Optional[str], stt_model: str = "small", tts_voice: str = "en_US-lessac-medium") -> Engines:
     root = Path(data_dir) if data_dir else Path.home() / ".friend" / "voice" / "models"
     root.mkdir(parents=True, exist_ok=True)
     speaker = None

package/dist/voice/friend_voice/stt.py

@@ -10,6 +10,17 @@ from typing import Optional
 MAX_SECONDS = 90.0
 SAMPLE_RATE = 16000
 
+#: Speech quieter than this (RMS) is brought up toward it; a speaker far from the microphone arrives faint.
+TARGET_RMS = 0.06
+#: Below this the room is silent, so there is nothing to bring up (it would only make hiss louder).
+SILENCE_RMS = 0.001
+MAX_GAIN = 20.0
+#: A segment is dropped when the model thinks it is probably not speech AND is not sure of its words (Whisper's own rule),
+#: or when it is very unsure of its words whatever the speech chance: noise turns into invented words in other scripts.
+NO_SPEECH_PROB = 0.6
+MIN_LOGPROB = -1.0
+HOPELESS_LOGPROB = -2.0
+
 #: A sentence in the style wanted, given to the recognizer as a hint: Hindi and Marathi are written in Devanagari,
 #: and the English words that people mix in stay in Latin letters.
 HINT_PROMPTS = {
@@ -18,6 +29,29 @@ HINT_PROMPTS = {
 }
 
 
+def level(samples):
+    """Bring faint speech up to a usable volume without clipping; leave loud enough speech and silence alone."""
+    import numpy as np
+
+    if samples.size == 0:
+        return samples
+    loudness = float(np.sqrt(np.mean(np.square(samples))))
+    if loudness >= TARGET_RMS or loudness < SILENCE_RMS:
+        return samples
+    peak = float(np.max(np.abs(samples)))
+    gain = min(TARGET_RMS / loudness, MAX_GAIN, 0.99 / peak if peak > 0 else MAX_GAIN)
+    return (samples * gain).astype(np.float32)
+
+
+def believed(segment) -> bool:
+    """Whether to keep what the model heard in this segment."""
+    no_speech = getattr(segment, "no_speech_prob", 0.0)
+    logprob = getattr(segment, "avg_logprob", 0.0)
+    if no_speech > NO_SPEECH_PROB and logprob < MIN_LOGPROB:
+        return False
+    return logprob >= HOPELESS_LOGPROB
+
+
 def decode_audio(data: bytes, max_seconds: float = MAX_SECONDS):
     """Decode any container PyAV understands (webm/opus, mp4/aac, wav, ...) to mono float32 at 16 kHz.
 
@@ -49,7 +83,7 @@ def decode_audio(data: bytes, max_seconds: float = MAX_SECONDS):
         take(resampler.resample(None))  # flush
     if not pieces:
         return np.zeros(0, dtype=np.float32)
-    return np.concatenate(pieces).astype(np.float32) / 32768.0
+    return level(np.concatenate(pieces).astype(np.float32) / 32768.0)
 
 
 class WhisperSTT:
@@ -64,12 +98,15 @@ class WhisperSTT:
         segments, info = self._model.transcribe(
             samples,
             language=language,
-            beam_size=1,
+            beam_size=5,  # a wider search is slower but much better on faint or accented speech
             vad_filter=True,  # drop silence so quiet clips return an empty transcript
             condition_on_previous_text=False,
+            no_speech_threshold=NO_SPEECH_PROB,
+            log_prob_threshold=MIN_LOGPROB,
+            compression_ratio_threshold=2.4,  # repeated gibberish compresses well: the model tries again instead of keeping it
             **extra,
         )
-        return " ".join(s.text.strip() for s in segments).strip(), info
+        return " ".join(s.text.strip() for s in segments if believed(s)).strip(), info
 
     def transcribe(self, audio: bytes, language: Optional[str]) -> dict:
         samples = decode_audio(audio)  # bounded decode: hostile or huge input is refused here