feat: tune Whisper settings and add word mappings

google7181fa1faffcf274.html

@@ -0,0 +1 @@
+google-site-verification: google7181fa1faffcf274.html

local_worker/vod_backfiller.py

@@ -37,6 +37,7 @@ TWITCH_CHANNEL = os.getenv("TWITCH_CHANNEL", "winx_prinx").lower()
 WHISPER_MODEL_SIZE = os.getenv("WHISPER_MODEL", "base")
 WHISPER_DEVICE = os.getenv("WHISPER_DEVICE", "cpu")
 WHISPER_COMPUTE_TYPE = os.getenv("WHISPER_COMPUTE_TYPE", "int8")
+WHISPER_CPU_THREADS = int(os.getenv("WHISPER_CPU_THREADS", "2"))
 
 if not API_KEY:
     print("[Error] API_KEY is missing in .env! Cannot push data to backend.")
@@ -257,10 +258,11 @@ def transcribe_vod_audio(vod_id, start_time_iso, model, start_offset=0, end_offs
                 beam_size=5, 
                 language="ru",
                 temperature=0.0,
+                condition_on_previous_text=False,
                 log_prob_threshold=-0.8,
                 no_speech_threshold=0.6,
                 vad_filter=True,
-                vad_parameters=dict(min_silence_duration_ms=500)
+                vad_parameters=dict(threshold=0.5, min_silence_duration_ms=500)
             )
             
             words_list = []
@@ -391,7 +393,8 @@ if __name__ == "__main__":
             model = WhisperModel(
                 WHISPER_MODEL_SIZE, 
                 device=WHISPER_DEVICE, 
-                compute_type=WHISPER_COMPUTE_TYPE
+                compute_type=WHISPER_COMPUTE_TYPE,
+                cpu_threads=WHISPER_CPU_THREADS
             )
             transcribe_vod_audio(vod_id, created_at, model)
         except Exception as e:

local_worker/worker.py

@@ -44,6 +44,7 @@ CAPTURE_METHOD = os.getenv("CAPTURE_METHOD", "stream").lower()
 WHISPER_MODEL_SIZE = os.getenv("WHISPER_MODEL", "base")
 WHISPER_DEVICE = os.getenv("WHISPER_DEVICE", "cpu")
 WHISPER_COMPUTE_TYPE = os.getenv("WHISPER_COMPUTE_TYPE", "int8")
+WHISPER_CPU_THREADS = int(os.getenv("WHISPER_CPU_THREADS", "2"))
 
 # Verify essential secrets
 if not API_KEY:
@@ -216,12 +217,13 @@ def chat_sender():
 
 def init_whisper_model():
     """Load Faster-Whisper model into memory"""
-    print(f"[Whisper] Loading model '{WHISPER_MODEL_SIZE}' on {WHISPER_DEVICE} ({WHISPER_COMPUTE_TYPE})...")
+    print(f"[Whisper] Loading model '{WHISPER_MODEL_SIZE}' on {WHISPER_DEVICE} ({WHISPER_COMPUTE_TYPE}, threads={WHISPER_CPU_THREADS})...")
     # This might take a few minutes on first run as the model downloads
     model = WhisperModel(
         WHISPER_MODEL_SIZE, 
         device=WHISPER_DEVICE, 
-        compute_type=WHISPER_COMPUTE_TYPE
+        compute_type=WHISPER_COMPUTE_TYPE,
+        cpu_threads=WHISPER_CPU_THREADS
     )
     print("[Whisper] Model loaded successfully.")
     return model
@@ -240,10 +242,11 @@ def transcribe_audio_segment(model, pcm_bytes):
         beam_size=5, 
         language="ru",
         temperature=0.0,
+        condition_on_previous_text=False,
         log_prob_threshold=-0.8,
         no_speech_threshold=0.6,
         vad_filter=True, # Voice Activity Detection filters out silence
-        vad_parameters=dict(min_silence_duration_ms=500)
+        vad_parameters=dict(threshold=0.5, min_silence_duration_ms=500)
     )
     
     words_list = []
@@ -516,10 +519,11 @@ def run_microphone_capture(model):
                     beam_size=5, 
                     language="ru",
                     temperature=0.0,
+                    condition_on_previous_text=False,
                     log_prob_threshold=-0.8,
                     no_speech_threshold=0.6,
                     vad_filter=True,
-                    vad_parameters=dict(min_silence_duration_ms=500)
+                    vad_parameters=dict(threshold=0.5, min_silence_duration_ms=500)
                 )
                 words_list = []
                 for segment in segments:

server/server.js

@@ -532,6 +532,18 @@ const STOP_WORDS = new Set([
   'here', 'whats', 'what', 'which', 'who', 'whom', 'whose', 'why', 'how', 'can', 'will', 'just'
 ]);
 
+// Word mappings/synonyms to group similar pronunciations or typos
+const WORD_MAPPINGS = {
+  'блядь': 'блять',
+  'петерка': 'пятерка',
+  'питерка': 'пятерка',
+  'петёрка': 'пятерка',
+  'пётерка': 'пятерка',
+  'кримбал': 'кримбр',
+  'кримбл': 'кримбр',
+  'кримбор': 'кримбр'
+};
+
 // Get word frequencies (Voice vs Chat messages from streamer)
 app.get('/api/stats/words', cacheMiddleware(), async (req, res) => {
   const streamId = req.query.stream_id ? parseInt(req.query.stream_id) : null;
@@ -546,8 +558,8 @@ app.get('/api/stats/words', cacheMiddleware(), async (req, res) => {
     for (const w of rawWords) {
       if (!w || !w.word || typeof w.word !== 'string') continue;
       let clean = w.word.toLowerCase().trim();
-      if (clean === 'блядь') {
-        clean = 'блять';
+      if (WORD_MAPPINGS[clean]) {
+        clean = WORD_MAPPINGS[clean];
       }
       
       const isAllowedLength = clean.length >= 3 || clean === 'ну';
@@ -602,8 +614,8 @@ app.get('/api/stats/words', cacheMiddleware(), async (req, res) => {
         }
         if (!word) continue;
 
-        if (word === 'блядь') {
-          word = 'блять';
+        if (WORD_MAPPINGS[word]) {
+          word = WORD_MAPPINGS[word];
         }
 
         const isAllowedLength = word.length >= 3 || word === 'ну';