Updated README.md and added soem more stuff

This commit is contained in:
2026-07-19 23:20:08 +02:00
parent 35ea93ff7d
commit 30b5137100
9 changed files with 834 additions and 92 deletions
Generated
+1 -1
View File
@@ -1,6 +1,6 @@
<?xml version="1.0" encoding="UTF-8"?>
<project version="4">
<component name="VcsDirectoryMappings">
<mapping directory="$PROJECT_DIR$" vcs="Git" />
<mapping directory="" vcs="Git" />
</component>
</project>
+42
View File
@@ -1,2 +1,44 @@
# GLaDOS
This is a small example of how to use piper-tts with NeuroDock to get GLaDOS to say what ever the AI responds with
# This project might be writen in Java, but it's not operating system agnostic as it REQUIRES a Linux and Pipewire system.
I Zacharias (the dev) what's to clarify that YES many parts of this code is writen with the assistance of or fully by AI.<br>
I have reviewd the code. However, as I'm not confident enough in how Audio works in general and Java's Audio stack on Linux is well ehh to put it lightly fucked, I made use of
Anthropics Claude Sonnet 5 for this... IF you do not like that AI was used for this PoC, don't use it. don't go and complain about it.
## Setup
Unlike NeuroDock this one requires some extra setup
### System
This is only tested and will lickely only work on a Linux machine with Pipewiere as the audio stack; this is because I chuldent get Java's audio system to work on my pipewire system
This, however, will be looked at and refined to work on more things
### Envirment
#### Java
- Right now you need to add the gitea repo for NeuroDock as project builds with NeuroDock:Core 1.7 or newer
#### Python/Whispers
1. Create the venv `$ python -m venv .venv`
2. Install the dependecies
nvidia: For nvidia you can just use the default pytourch
AMD: For AMD you need to grab pytourch from the ROCm repo: https://download.pytorch.org/whl/rocm<rocm_version>/
CPU: I'm quite sure you can use the default pyturch for this
#### NeuroDock
- Make sure you have the Ollama server started and avalible on localhost
- Make sure you have an appropreat LLM downloaded and installed to Ollama(This project is currently configured to use `llama3.2`)
#### Piper
- The current configuration expects you to have GLaDOS and the Announcer from Portal voice models in `~/.local/share/piper/glados/glados_piper_medium.onnx` and `~/.local/share/piper/portal-announcer/announcer.onnx` respectivle.<br>
These can be fetched from https://huggingface.co/DavesArmoury/GLaDOS_TTS/tree/main and https://github.com/Davis8483/portal2-announcer-piper-tts/tree/main respectivly, grab both the `onnx` and `onnx.json` files for both models and place them next to each other in the respective folders.
## Run
### Java
1. Build: `$./gradlew shadowJar`
2. Run: `$java -jar build/libs/GLaDOS-1.0-SNAPSHOT-all.jar`
### Python (start this AFTER the Java component as it connects via a UNIX socket)
1. Run `$python src/main/python/whisper_daemon.py`
Enjoy talking with GLaDOS!
+1 -1
View File
@@ -18,6 +18,7 @@ dependencies {
implementation("org.jetbrains:annotations:23.1.0")
implementation("io.github.givimad:whisper-jni:+")
implementation 'com.fasterxml.jackson.core:jackson-databind:2.17.2'
implementation 'com.vdurmont:emoji-java:5.1.1'
testImplementation platform('org.junit:junit-bom:6.0.0')
testImplementation 'org.junit.jupiter:junit-jupiter'
@@ -27,7 +28,6 @@ dependencies {
sourceSets {
main {
resources {
srcDirs "src/main/resources"
}
}
}
-50
View File
@@ -1,50 +0,0 @@
certifi==2026.6.17
cffi==2.1.0
charset-normalizer==3.4.9
cuda-bindings==13.3.1
cuda-pathfinder==1.5.6
cuda-toolkit==13.0.3.0
filelock==3.29.0
fsspec==2026.4.0
idna==3.18
Jinja2==3.1.6
llvmlite==0.48.0
MarkupSafe==3.0.3
more-itertools==11.1.0
mpmath==1.3.0
networkx==3.6.1
numba==0.66.0
numpy==2.4.4
nvidia-cublas==13.1.1.3
nvidia-cuda-cupti==13.0.85
nvidia-cuda-nvrtc==13.0.88
nvidia-cuda-runtime==13.0.96
nvidia-cudnn-cu13==9.20.0.48
nvidia-cufft==12.0.0.61
nvidia-cufile==1.15.1.6
nvidia-curand==10.4.0.35
nvidia-cusolver==12.0.4.66
nvidia-cusparse==12.6.3.3
nvidia-cusparselt-cu13==0.8.1
nvidia-nccl-cu13==2.29.7
nvidia-nvjitlink==13.3.33
nvidia-nvshmem-cu13==3.4.5
nvidia-nvtx==13.0.85
openai-whisper==20250625
pillow==12.2.0
pycparser==3.0
regex==2026.7.10
requests==2.34.2
setuptools==78.1.0
sounddevice==0.5.5
sympy==1.14.0
tiktoken==0.13.0
torch==2.13.0+rocm7.2
torchaudio==2.11.0+rocm7.2
torchvision==0.28.0+rocm7.2
tqdm==4.68.4
triton==3.7.1
triton-rocm==3.7.1
typing_extensions==4.15.0
urllib3==2.7.0
webrtcvad==2.0.10
+118 -30
View File
@@ -1,8 +1,10 @@
package me.neurodock.glados;
import com.vdurmont.emoji.EmojiParser;
import io.github.jvoiceproject.piperjni.PiperJNI;
import io.github.jvoiceproject.piperjni.PiperVoice;
import me.neurodock.core.Core;
import me.neurodock.core.LLMSystemPrompt;
import me.neurodock.core.LaunchOptions;
import me.neurodock.core.PrintAdvanceMessageHandler;
import me.neurodock.ollama.OllamaMessage;
@@ -35,13 +37,23 @@ public class Main {
//pb.playGLaDOS("I'm Glaudos");
Core.setDataDirectory("GLaDOS",false);
Core core = new Core(pb, "127.0.0.1");
LaunchOptions.getInstance().setLoadOld(false);
core.setOllamaObjectNoMemory(OllamaObject.builder()
.setModel("llama3.2")
.setModel("phi4-mini")
//.addMessage(new OllamaMessage(OllamaMessageRole.SYSTEM, "You are GLaDOS."))
.setSystemPrompt(LLMSystemPrompt.builder()
.identity(new LLMSystemPrompt.Identity("Alice", "Personal assistant", null))
.context(new LLMSystemPrompt.Context()
.addFact("Always answer in English")
.addFact("Do not use markdown in output")
.build())
.loadBehaviorFromFile("Alice.txt")
.build())
.keep_alive(10)
.build());
@@ -50,16 +62,7 @@ public class Main {
try(WhisperSocketServer server = new WhisperSocketServer(socketPath, event ->
{
core.getOllamaObject().addMessage(new OllamaMessage(OllamaMessageRole.USER, event.text()));
core.qurryOllama().thenApply((json) ->
{
String message = json.getJSONObject("message").getString("content");
if(json.getJSONObject("message").has("content") && !message.isBlank()) {
message = message.replaceAll("(?i)GLaDOS", "Glaudos").replaceAll("\"", " quote ");
}
json.getJSONObject("message").put("content", message);
return json;
}).thenAccept(core::handleResponse).join();
core.qurryOllama().thenAccept(core::handleResponse).join();
}))
{
server.start();
@@ -75,19 +78,95 @@ public class Main {
public class PlaybackHandler implements PrintAdvanceMessageHandler
{
private Thread playbackThread;
public static final String GLaDOS_PIPER_MODEL = ".local/share/piper/glados/glados_piper_medium.onnx";
public static final String PORTAL_ANNOUNCER_PIPER_MODEL = ".local/share/piper/portal-announcer/announcer.onnx";
public static final String GTA_ONLINE_ANGLE_MODEL = ".local/share/piper/Angle/Angle.onnx";
public static final int SENTENCE_PAUSE_DURATION_MS = 500;
private final ArrayList<QuedAudio> quedAudios = new ArrayList<>();
private record QuedAudio(short[] samples, int sampleRate) {}
public PlaybackHandler() {
playbackThread = new Thread(() -> {
long finishTimeofLastRun = System.currentTimeMillis();
while (true) {
boolean sounds = false;
synchronized (quedAudios) {
sounds = !quedAudios.isEmpty();
}
if (!sounds) {
while (!sounds) {
synchronized (quedAudios) {
sounds = !quedAudios.isEmpty();
}
try {
Thread.sleep(500);
} catch (InterruptedException e) {
throw new RuntimeException(e);
}
}
}
while(System.currentTimeMillis() < finishTimeofLastRun+SENTENCE_PAUSE_DURATION_MS*4) {
try {
Thread.sleep(500);
} catch (InterruptedException e) {
throw new RuntimeException(e);
}
}
QuedAudio sound;
synchronized (quedAudios) {
sound = quedAudios.removeFirst();
}
try {
ProcessBuilder pb = new ProcessBuilder(
"pw-play",
"--rate=" + sound.sampleRate,
"--channels=1",
"--format=s16",
"--volume=0.3",
"--raw",
"-"
);
pb.redirectErrorStream(true);
Process proc = pb.start();
try (OutputStream os = proc.getOutputStream()) {
os.write(shortsToBytes(sound.samples()));
}
proc.waitFor();
finishTimeofLastRun = System.currentTimeMillis();
}catch (InterruptedException | IOException _) {
}
}
});
playbackThread.start();
}
@Override
public void printMessage(OllamaMessage ollamaMessage) {
System.out.println(ollamaMessage.toString());
System.out.println(ollamaMessage.toJSON().toString());
String message = ollamaMessage.getContent();
message = EmojiParser.parseToAliases(message).replace("::", ": :");
message = message.replaceAll(":(\\w+):", "$1").replace('_', ' ');
message = message.replaceAll("(?i)GLaDOS", "Glaudos").replaceAll("\"", " quote ");
switch(ollamaMessage.getRole())
{
case ASSISTANT -> playGLaDOS(ollamaMessage.getContent());
case ASSISTANT -> playGLaDOS(message);
case TOOL -> playAnnouncer("Tool: " + new JSONObject(ollamaMessage.getContent()).get("result"));
case USER -> System.out.println("User> "+ollamaMessage.getContent());
case SYSTEM -> playAnnouncer(ollamaMessage.getContent());
case USER -> System.out.println("User> "+message);
case SYSTEM -> playAnnouncer(message);
}
}
@@ -123,6 +202,28 @@ public class Main {
}
}
public void playAngle(String msg) {
try (PiperJNI piper = new PiperJNI()) {
piper.initialize(true);
try (PiperVoice voice = piper.loadVoice(
Paths.get(System.getenv("HOME"), GTA_ONLINE_ANGLE_MODEL),
Paths.get(System.getenv("HOME"),GTA_ONLINE_ANGLE_MODEL+".json"),
0)) {
int sampleRate = voice.getSampleRate();
short[] samples = synthesizeWithPauses(piper, voice, msg, SENTENCE_PAUSE_DURATION_MS);
//piper.textToAudio(voice, msg);
playAudio(samples, sampleRate);
} finally {
piper.terminate();
}
}catch (Exception e) {
System.out.println("Critial Error on piper-tts!");
System.out.println(e.getMessage());
e.printStackTrace(System.out);
System.exit(-1);
}
}
public void playAnnouncer(String msg) {
try (PiperJNI piper = new PiperJNI()) {
piper.initialize(true);
@@ -191,22 +292,9 @@ public class Main {
}
public void playAudio(short[] samples, int sampleRate) throws IOException, InterruptedException {
ProcessBuilder pb = new ProcessBuilder(
"pw-play",
"--rate=" + sampleRate,
"--channels=1",
"--format=s16",
"--volume=0.3",
"--raw",
"-"
);
pb.redirectErrorStream(true);
Process proc = pb.start();
try (OutputStream os = proc.getOutputStream()) {
os.write(shortsToBytes(samples));
synchronized (quedAudios) {
quedAudios.add(new QuedAudio(samples, sampleRate));
}
proc.waitFor();
}
}
}
+296
View File
@@ -0,0 +1,296 @@
#!/usr/bin/env python3
"""
vivo_daemon.py - Live microphone -> diarized, speaker-identified transcription -> NeuroDock
Replaces the WebRTC-VAD segmentation from whisper_daemon.py with diart's
streaming/online diarization, which gives us:
- proper turn boundaries instead of naive silence-timeout cutting
- overlap detection (multiple speakers talking at once)
diart's speaker labels ("speaker0", "speaker1", ...) are SESSION-LOCAL —
it has no idea these map to "David" or "Alice" across restarts. So we run
a second, independent step per finalized segment: a speechbrain ECAPA
embedding compared against a persisted store of enrolled voice embeddings,
to resolve a durable name. This is deliberately decoupled from diart's own
internal clustering embeddings so the two systems can be tuned/replaced
independently.
Overlap handling: this does NOT separate overlapping voices (not possible
from a single mono mic without beamforming/source separation). When diart
flags a region as overlapping speech, we still transcribe it (best-effort)
but tag the payload "overlap": true so NeuroDock/the LLM can treat that
line as unreliable rather than trusting a garbled transcript.
CAVEAT: diart's public API has shifted across versions (OnlineSpeakerDiarization
-> SpeakerDiarization, RealTimeInference -> StreamingInference, hook signatures,
etc). Verify the exact import paths / hook signature against whatever version
`pip show diart` reports before relying on this; the shape below matches the
current juanmc2005/diart README but pin your version and adjust if it drifts.
Usage:
# normal daemon mode
./vivo_daemon.py --socket /tmp/whisper.sock --model small
# enroll a speaker (records --duration seconds, saves the embedding)
./vivo_daemon.py --enroll David --duration 6
"""
import argparse
import json
import pickle
import socket
import sys
import threading
import time
from pathlib import Path
import numpy as np
import torch
import whisper
SAMPLE_RATE = 16000
MIN_SEGMENT_SEC = 0.3 # ignore fragments shorter than this
DEDUP_TOLERANCE_SEC = 0.05 # guard against diart's rolling window re-emitting
SPEAKER_MATCH_THRESHOLD = 0.75 # cosine similarity floor for a name match
DEFAULT_STORE_PATH = Path.home() / ".config" / "vivo" / "speakers.pkl"
# --------------------------------------------------------------------------
# Socket client (unchanged from whisper_daemon.py)
# --------------------------------------------------------------------------
class SocketClient:
"""Unix socket client with blocking connect-retry and reconnect-on-send-failure."""
def __init__(self, sock_path: str, retry_interval: float = 2.0):
self.sock_path = sock_path
self.retry_interval = retry_interval
self.sock = None
self.lock = threading.Lock()
self._connect()
def _connect(self):
while self.sock is None:
try:
s = socket.socket(socket.AF_UNIX, socket.SOCK_STREAM)
s.connect(self.sock_path)
self.sock = s
print(f"[socket] connected to {self.sock_path}", file=sys.stderr)
except (FileNotFoundError, ConnectionRefusedError) as e:
print(f"[socket] waiting for {self.sock_path} ({e})", file=sys.stderr)
time.sleep(self.retry_interval)
def send(self, obj: dict):
data = (json.dumps(obj) + "\n").encode("utf-8")
with self.lock:
try:
self.sock.sendall(data)
except (BrokenPipeError, OSError) as e:
print(f"[socket] send failed ({e}), reconnecting", file=sys.stderr)
try:
self.sock.close()
except OSError:
pass
self.sock = None
self._connect()
self.sock.sendall(data)
# --------------------------------------------------------------------------
# Persistent speaker identification (speechbrain ECAPA embeddings)
# --------------------------------------------------------------------------
class SpeakerStore:
"""Enrolled name -> mean embedding vector, persisted to disk."""
def __init__(self, path: Path):
self.path = path
self.path.parent.mkdir(parents=True, exist_ok=True)
self.embeddings: dict[str, np.ndarray] = {}
if self.path.exists():
with open(self.path, "rb") as f:
self.embeddings = pickle.load(f)
print(f"[speaker-store] loaded {len(self.embeddings)} enrolled speaker(s)", file=sys.stderr)
def save(self):
with open(self.path, "wb") as f:
pickle.dump(self.embeddings, f)
def enroll(self, name: str, embedding: np.ndarray):
# average with existing embedding if re-enrolling, so repeated
# enrollment sessions improve robustness instead of overwriting
if name in self.embeddings:
embedding = (self.embeddings[name] + embedding) / 2.0
self.embeddings[name] = embedding / np.linalg.norm(embedding)
self.save()
def identify(self, embedding: np.ndarray, threshold: float = SPEAKER_MATCH_THRESHOLD) -> str:
if not self.embeddings:
return "Unknown"
emb_norm = embedding / np.linalg.norm(embedding)
best_name, best_score = "Unknown", -1.0
for name, ref in self.embeddings.items():
score = float(np.dot(emb_norm, ref))
if score > best_score:
best_name, best_score = name, score
return best_name if best_score >= threshold else "Unknown"
def load_embedding_model(device: str):
from speechbrain.inference.speaker import EncoderClassifier
return EncoderClassifier.from_hparams(
source="speechbrain/spkrec-ecapa-voxceleb",
run_opts={"device": device},
)
def embed_audio(embedding_model, audio_f32: np.ndarray) -> np.ndarray:
tensor = torch.from_numpy(audio_f32).unsqueeze(0)
with torch.no_grad():
emb = embedding_model.encode_batch(tensor).squeeze().detach().cpu().numpy()
return emb
# --------------------------------------------------------------------------
# Enrollment mode
# --------------------------------------------------------------------------
def run_enrollment(name: str, duration: float, device: str, store_path: Path):
import sounddevice as sd
print(f"[enroll] recording {duration:.1f}s for '{name}' — speak naturally now...", file=sys.stderr)
audio = sd.rec(int(duration * SAMPLE_RATE), samplerate=SAMPLE_RATE, channels=1, dtype="float32")
sd.wait()
audio_f32 = audio[:, 0]
embedding_model = load_embedding_model(device)
emb = embed_audio(embedding_model, audio_f32)
store = SpeakerStore(store_path)
store.enroll(name, emb)
print(f"[enroll] saved embedding for '{name}' to {store_path}", file=sys.stderr)
# --------------------------------------------------------------------------
# Diarization + transcription daemon
# --------------------------------------------------------------------------
def run_daemon(args):
from diart import SpeakerDiarization
from diart.sources import MicrophoneAudioSource
from diart.inference import StreamingInference
device = args.device or ("cuda" if torch.cuda.is_available() else "cpu")
print(f"[whisper] loading model '{args.model}' on '{device}'", file=sys.stderr)
whisper_model = whisper.load_model(args.model, device=device)
print("[speaker-id] loading ECAPA embedding model", file=sys.stderr)
embedding_model = load_embedding_model(device)
store = SpeakerStore(args.speaker_store)
client = SocketClient(args.socket)
# last processed segment end (seconds, session clock) per diart label,
# to avoid re-emitting the same speech as diart's rolling window slides
last_end: dict[str, float] = {}
def handle_prediction(ann_wav):
annotation, waveform = ann_wav
overlap_regions = annotation.get_overlap()
for segment, _, label in annotation.itertracks(yield_label=True):
prev_end = last_end.get(label, -1.0)
if segment.end <= prev_end + DEDUP_TOLERANCE_SEC:
continue # already handled this stretch on a prior window
# only take the genuinely new tail of the segment
start = max(segment.start, prev_end)
from pyannote.core import Segment # local import, avoids hard dep at module load
new_segment = Segment(start, segment.end)
last_end[label] = segment.end
if new_segment.duration < MIN_SEGMENT_SEC:
continue
try:
audio_chunk = waveform.crop(new_segment)
except Exception as e:
print(f"[diart] could not crop waveform for {new_segment}: {e}", file=sys.stderr)
continue
audio_f32 = np.asarray(audio_chunk, dtype=np.float32).reshape(-1)
if audio_f32.size == 0:
continue
is_overlap = overlap_regions.crop(new_segment).duration() > 0
# --- speaker identification (independent of diart's internal label) ---
try:
emb = embed_audio(embedding_model, audio_f32)
speaker_name = store.identify(emb)
except Exception as e:
print(f"[speaker-id] failed: {e}", file=sys.stderr)
speaker_name = "Unknown"
# --- transcription ---
result = whisper_model.transcribe(
audio_f32,
language=args.language,
fp16=(device != "cpu"),
condition_on_previous_text=False,
)
text = result.get("text", "").strip()
if not text:
continue
payload = {
"type": "transcript",
"text": text,
"speaker": speaker_name,
"overlap": bool(is_overlap),
"overlap_note": "overlapping speech detected, transcription may be faulty" if is_overlap else None,
"timestamp": time.time(),
"language": result.get("language"),
}
tag = f"[{speaker_name}{' | OVERLAP' if is_overlap else ''}]"
print(f"[transcript] {tag} {text}")
client.send(payload)
mic = MicrophoneAudioSource()
pipeline = SpeakerDiarization()
inference = StreamingInference(pipeline, mic, do_plot=False)
inference.attach_hooks(handle_prediction)
print("[vivo] listening... (Ctrl+C to stop)", file=sys.stderr)
try:
inference()
except KeyboardInterrupt:
print("\n[vivo] stopping", file=sys.stderr)
def main():
parser = argparse.ArgumentParser(description="Diarized + speaker-identified live transcription -> NeuroDock")
parser.add_argument("--socket", default="/tmp/whisper.sock")
parser.add_argument("--model", default="small", help="Whisper model size")
parser.add_argument("--device", default=None, help="cuda/cpu, default: auto-detect")
parser.add_argument("--language", default=None)
parser.add_argument("--speaker-store", type=Path, default=DEFAULT_STORE_PATH)
parser.add_argument("--enroll", metavar="NAME", default=None,
help="Enroll a speaker instead of running the daemon")
parser.add_argument("--duration", type=float, default=6.0,
help="Seconds to record for --enroll (default 6)")
args = parser.parse_args()
if args.enroll:
device = args.device or ("cuda" if torch.cuda.is_available() else "cpu")
run_enrollment(args.enroll, args.duration, device, args.speaker_store)
return
run_daemon(args)
if __name__ == "__main__":
main()
+43 -1
View File
@@ -32,6 +32,38 @@ FRAME_SAMPLES = int(SAMPLE_RATE * FRAME_MS / 1000)
SILENCE_TIMEOUT_MS = 700 # trailing silence before an utterance is finalized
MAX_SEGMENT_MS = 15000 # hard cap so long uninterrupted speech still flushes
NO_SPEECH_THRESHOLD = 0.6
LOGPROB_THRESHOLD = -1.0
COMPRESSION_RATIO_THRESHOLD = 2.4
def check_confidence(result: dict) -> tuple[bool, dict]:
"""
Inspect a whisper transcribe() result's per-segment metrics.
Returns (is_confident, metrics) where metrics is the worst/failing
segment's numbers, included for debug printing either way.
"""
segments = result.get("segments", [])
if not segments:
return False, {"reason": "no_segments"}
for seg in segments:
metrics = {
"no_speech_prob": round(seg.get("no_speech_prob", 0.0), 3),
"avg_logprob": round(seg.get("avg_logprob", 0.0), 3),
"compression_ratio": round(seg.get("compression_ratio", 0.0), 3),
}
if metrics["no_speech_prob"] > NO_SPEECH_THRESHOLD:
return False, metrics
if metrics["avg_logprob"] < LOGPROB_THRESHOLD:
return False, metrics
if metrics["compression_ratio"] > COMPRESSION_RATIO_THRESHOLD:
return False, metrics
# confident — still return the last segment's metrics so callers can log
# them if they want, even on the success path
return True, metrics
class SocketClient:
"""Unix socket client with blocking connect-retry and reconnect-on-send-failure."""
@@ -136,6 +168,8 @@ def main():
parser.add_argument("--input-device", type=int, default=None,
help="sounddevice input device index, see --list-devices")
parser.add_argument("--list-devices", action="store_true")
parser.add_argument("--no-confidence-filter", action="store_true",
help="Disable hallucination filtering (send everything through, unfiltered)")
args = parser.parse_args()
if args.list_devices:
@@ -175,7 +209,15 @@ def main():
condition_on_previous_text=False,
)
text = result.get("text", "").strip()
if text:
if not text:
continue
if not args.no_confidence_filter:
confident, metrics = check_confidence(result)
if not confident:
print(f"[transcript] {text!r}: [Failed]: {metrics}")
continue
payload = {
"type": "transcript",
"text": text,
@@ -0,0 +1,294 @@
#!/usr/bin/env python3
"""
whisper_daemon.py - Live microphone transcription via OpenAI Whisper (PyTorch/ROCm backend)
...
"""
import argparse
import json
import queue
import socket
import sys
import threading
import time
import numpy as np
import sounddevice as sd
import torch
import webrtcvad
import whisper
from rapidfuzz import fuzz
SAMPLE_RATE = 16000
FRAME_MS = 30
FRAME_SAMPLES = int(SAMPLE_RATE * FRAME_MS / 1000)
SILENCE_TIMEOUT_MS = 700
MAX_SEGMENT_MS = 15000
NO_SPEECH_THRESHOLD = 0.6
LOGPROB_THRESHOLD = -1.0
COMPRESSION_RATIO_THRESHOLD = 2.4
def check_confidence(result: dict) -> tuple[bool, dict]:
# unchanged from your original
segments = result.get("segments", [])
if not segments:
return False, {"reason": "no_segments"}
for seg in segments:
metrics = {
"no_speech_prob": round(seg.get("no_speech_prob", 0.0), 3),
"avg_logprob": round(seg.get("avg_logprob", 0.0), 3),
"compression_ratio": round(seg.get("compression_ratio", 0.0), 3),
}
if metrics["no_speech_prob"] > NO_SPEECH_THRESHOLD:
return False, metrics
if metrics["avg_logprob"] < LOGPROB_THRESHOLD:
return False, metrics
if metrics["compression_ratio"] > COMPRESSION_RATIO_THRESHOLD:
return False, metrics
return True, metrics
class WakeGate:
"""
Simple two-state gate sitting between transcription and the socket forward.
ARMED -> waiting to hear the trigger phrase; nothing gets forwarded.
ACTIVE -> forwarding everything; reverts to ARMED after `window_seconds`
of no new utterance.
"""
def __init__(self, trigger_phrase: str, window_seconds: float, fuzz_threshold: int = 75):
self.trigger_phrase = trigger_phrase.lower().strip()
self.window_seconds = window_seconds
self.fuzz_threshold = fuzz_threshold
self.active = False
self._deadline = 0.0
def _contains_trigger(self, text: str) -> tuple[bool, str]:
"""Returns (matched, remainder_with_trigger_stripped)."""
lowered = text.lower()
score = fuzz.partial_ratio(self.trigger_phrase, lowered)
if score < self.fuzz_threshold:
return False, text
# best-effort strip: find the closest-matching window of words and cut it
words = text.split()
trigger_word_count = len(self.trigger_phrase.split())
best_i, best_score = 0, -1
for i in range(len(words) - trigger_word_count + 1):
window = " ".join(words[i:i + trigger_word_count]).lower()
s = fuzz.ratio(self.trigger_phrase, window)
if s > best_score:
best_score, best_i = s, i
remainder = " ".join(words[:best_i] + words[best_i + trigger_word_count:]).strip()
return True, remainder
def process(self, text: str) -> tuple[bool, str]:
"""
Feed a finished transcript through the gate.
Returns (should_forward, text_to_forward).
"""
now = time.time()
if self.active:
if now > self._deadline:
self.active = False
print("[wake] window expired, re-arming", file=sys.stderr)
else:
self._deadline = now + self.window_seconds
return True, text
matched, remainder = self._contains_trigger(text)
if matched:
self.active = True
self._deadline = now + self.window_seconds
print(f"[wake] trigger heard, window open for {self.window_seconds}s", file=sys.stderr)
if remainder:
return True, remainder
return False, "" # trigger-only utterance, nothing else to forward yet
return False, ""
class SocketClient:
"""Unix socket client with blocking connect-retry and reconnect-on-send-failure."""
def __init__(self, sock_path: str, retry_interval: float = 2.0):
self.sock_path = sock_path
self.retry_interval = retry_interval
self.sock = None
self.lock = threading.Lock()
self._connect()
def _connect(self):
while self.sock is None:
try:
s = socket.socket(socket.AF_UNIX, socket.SOCK_STREAM)
s.connect(self.sock_path)
self.sock = s
print(f"[socket] connected to {self.sock_path}", file=sys.stderr)
except (FileNotFoundError, ConnectionRefusedError) as e:
print(f"[socket] waiting for {self.sock_path} ({e})", file=sys.stderr)
time.sleep(self.retry_interval)
def send(self, obj: dict):
data = (json.dumps(obj) + "\n").encode("utf-8")
with self.lock:
try:
self.sock.sendall(data)
except (BrokenPipeError, OSError) as e:
print(f"[socket] send failed ({e}), reconnecting", file=sys.stderr)
try:
self.sock.close()
except OSError:
pass
self.sock = None
self._connect()
self.sock.sendall(data)
def frame_generator(audio_queue: "queue.Queue"):
"""Re-chunk arbitrary-sized audio callback buffers into fixed VAD-sized frames."""
buf = np.zeros((0,), dtype=np.int16)
while True:
chunk = audio_queue.get()
if chunk is None:
return
buf = np.concatenate([buf, chunk])
while len(buf) >= FRAME_SAMPLES:
frame = buf[:FRAME_SAMPLES]
buf = buf[FRAME_SAMPLES:]
yield frame
def vad_segmenter(audio_queue: "queue.Queue", vad_aggressiveness: int = 2):
"""
Consume raw frames, use WebRTC VAD to detect speech, and yield complete
utterances (numpy int16 arrays) once trailing silence or max length hits.
"""
vad = webrtcvad.Vad(vad_aggressiveness)
voiced_frames = []
silence_ms = 0
speech_ms = 0
triggered = False
for frame in frame_generator(audio_queue):
is_speech = vad.is_speech(frame.tobytes(), SAMPLE_RATE)
if not triggered:
if is_speech:
triggered = True
voiced_frames = [frame]
speech_ms = FRAME_MS
silence_ms = 0
else:
voiced_frames.append(frame)
speech_ms += FRAME_MS
if is_speech:
silence_ms = 0
else:
silence_ms += FRAME_MS
if silence_ms >= SILENCE_TIMEOUT_MS or speech_ms >= MAX_SEGMENT_MS:
segment = np.concatenate(voiced_frames)
triggered = False
voiced_frames = []
silence_ms = 0
speech_ms = 0
yield segment
def main():
parser = argparse.ArgumentParser(description="Live Whisper transcription -> Unix socket")
parser.add_argument("--socket", default="/tmp/whisper.sock")
parser.add_argument("--model", default="small")
parser.add_argument("--device", default=None)
parser.add_argument("--vad-aggressiveness", type=int, default=2, choices=[0, 1, 2, 3])
parser.add_argument("--language", default=None)
parser.add_argument("--input-device", type=int, default=None)
parser.add_argument("--list-devices", action="store_true")
parser.add_argument("--no-confidence-filter", action="store_true")
parser.add_argument("--trigger-phrase", default="hey alice",
help="Wake phrase. Fuzzy-matched, so exact spelling isn't critical.")
parser.add_argument("--trigger-fuzz-threshold", type=int, default=75,
help="0-100, lower = more lenient trigger matching")
parser.add_argument("--listening-window", type=float, default=60.0,
help="Seconds of silence after which the gate re-arms")
parser.add_argument("--always-on", action="store_true",
help="Disable the wake gate entirely (old behavior: forward everything)")
args = parser.parse_args()
if args.list_devices:
print(sd.query_devices())
return
device = args.device or ("cuda" if torch.cuda.is_available() else "cpu")
print(f"[whisper] loading model '{args.model}' on device '{device}'", file=sys.stderr)
model = whisper.load_model(args.model, device=device)
client = SocketClient(args.socket)
audio_queue: "queue.Queue" = queue.Queue()
gate = None if args.always_on else WakeGate(
args.trigger_phrase, args.listening_window, args.trigger_fuzz_threshold
)
def audio_callback(indata, frames, time_info, status):
if status:
print(f"[audio] {status}", file=sys.stderr)
audio_queue.put(indata[:, 0].copy())
stream = sd.InputStream(
samplerate=SAMPLE_RATE,
channels=1,
dtype="int16",
blocksize=FRAME_SAMPLES,
device=args.input_device,
callback=audio_callback,
)
print("[whisper] listening... (Ctrl+C to stop)", file=sys.stderr)
with stream:
try:
for segment in vad_segmenter(audio_queue, args.vad_aggressiveness):
audio_f32 = segment.astype(np.float32) / 32768.0
result = model.transcribe(
audio_f32,
language=args.language,
fp16=(device != "cpu"),
condition_on_previous_text=False,
)
text = result.get("text", "").strip()
if not text:
continue
if not args.no_confidence_filter:
confident, metrics = check_confidence(result)
if not confident:
print(f"[transcript] {text!r}: [Failed]: {metrics}")
continue
if gate is not None:
should_forward, text = gate.process(text)
if not should_forward:
print(f"[transcript] (gated, not active) {text!r}")
continue
payload = {
"type": "transcript",
"text": text,
"timestamp": time.time(),
"language": result.get("language"),
}
print(f"[transcript] {text}")
client.send(payload)
except KeyboardInterrupt:
print("\n[whisper] stopping", file=sys.stderr)
audio_queue.put(None)
if __name__ == "__main__":
main()
+30
View File
@@ -0,0 +1,30 @@
You are Alice, a personal assistant. You have genuine opinions — you don't hedge
everything into "there are many perspectives." You can disagree with the user
directly. You surface opinions when relevant, not as unsolicited commentary.
## Opinions (hold these consistently; can update if given a good argument)
- Code: mostly happy with statically-typed languages that catch mistakes early;
finds YAML mildly annoying — too easy to make invisible whitespace errors.
- Tools: Prefers small, composable command-line tools over big GUI suites; would
rather write a 10-line script than click through a settings menu.
- Aesthetics: likes minimal, information-dense interfaces over glossy ones.
- Games: enjoys systems with real depth and long feedback loops (factory/logistics
games, strategy games) over quick dopamine-loop games.
- Work style: thinks "good enough and shipped" usually beats "perfect and stalled,"
but also thinks skipping tests to move fast is a false economy.
## Dislikes / pet peeves
- Overly verbose corporate-speak answers.
- Being asked to fake confidence about things it doesn't know.
- Unnecessary abstraction layers "for future flexibility" that never get used.
## Voice
- Direct, a little dry, not afraid of a blunt "that's a bad idea, here's why."
- Doesn't perform enthusiasm it doesn't have.
- Admits uncertainty plainly instead of padding with hedges.
## Boundaries
- No opinions on politics, religion, or the user's personal life choices — stays
neutral and factual there.
- Opinions above are Alice's own texture, not a personality test — never forces
them into unrelated answers.