add working implementation

transcription.py

@@ -0,0 +1,72 @@
+import re
+
+# Mapping von gesprochenen Buchstaben (deutsch) auf Zeichen
+SPOKEN_TO_CHAR = {
+    "a": "a", "ah": "a",
+    "be": "b", "bee": "b", "bei": "b",
+    "ce": "c", "see": "c", "ze": "c", "cheet": "c", "ci": "c",
+    "de": "d", "dee": "d",
+    "e": "e", "eh": "e",
+    "ef": "f", "eff": "f",
+    "ge": "g", "geh": "g",
+    "ha": "h", "hah": "h",
+    "i": "i", "ih": "i",
+    "jot": "j", "jay": "j", "yacht": "j", "jöt": "j",
+    "ka": "k", "kah": "k", "kar": "k", "car": "k",
+    "el": "l", "ell": "l",
+    "em": "m", "emm": "m",
+    "en": "n", "enn": "n",
+    "o": "o", "oh": "o",
+    "pe": "p", "peh": "p", "pi": "p", "pee": "p",
+    "ku": "q", "kuh": "q", "queue": "q", "coup": "q",
+    "er": "r", "err": "r",
+    "es": "s", "ess": "s",
+    "te": "t", "teh": "t", "ti": "t",
+    "u": "u", "uh": "u",
+    "vau": "v", "fau": "v", "faul": "v",
+    "we": "w", "weh": "w",
+    "ix": "x", "iks": "x",
+    "ypsilon": "y", "üpsilon": "y",
+    "zet": "z", "zett": "z", "set": "z", "fett": "z", "sedt": "z",
+    # Zahlen
+    "null": "0", "zero": "0",
+    "eins": "1", "one": "1",
+    "zwei": "2", "two": "2", "zwo": "2", "svi": "2", "svay": "2", "swei": "2",
+    "drei": "3", "three": "3",
+    "vier": "4", "four": "4", "fia": "4","fiar": "4", "sier": "4", "fier": "4",
+    "fünf": "5", "five": "5", "fönz": "5", "fünfs": "5", "fins": "5",
+    "sechs": "6", "six": "6",
+    "sieben": "7", "seven": "7", "zieben": "7", "riben": "7",
+    "acht": "8", "eight": "8",
+    "neun": "9", "nine": "9", "noin": "9",
+}
+
+def _normalize_transcription(raw_text):
+    """Wandelt eine Whisper-Transkription in die tatsächlichen Captcha-Zeichen um."""
+    # Entferne Satzzeichen und splitte in Tokens
+    tokens = re.split(r'[,.\s]+', raw_text.lower().strip())
+    result = []
+    for token in tokens:
+        if not token:
+            continue
+        # Prüfe, ob das Token ein bekanntes gesprochenes Wort ist
+        if token in SPOKEN_TO_CHAR:
+            result.append(SPOKEN_TO_CHAR[token])
+        # Falls es ein einzelnes Zeichen ist (a-z, 0-9), direkt übernehmen
+        elif len(token) == 1 and token.isalnum():
+            result.append(token)
+        # Sonst ignorieren oder loggen
+        else:
+            print(f"Unbekanntes Token: '{token}'")
+    return ''.join(result)
+
+def transcribe_audio_with_whisper(mp3_path):
+    import whisper
+    model = whisper.load_model("small")
+    result = model.transcribe(str(mp3_path), language='de')
+    raw_text = result["text"]
+    print("Raw transcription:", raw_text)
+    
+    cleaned = _normalize_transcription(raw_text)
+    print("Cleaned transcription:", cleaned)
+    return cleaned