Faster-Whisper-Leitfaden: Schnellere Sprach-zu-Text-Transkription mit CTranslate2

---

Was ist Faster-Whisper?

Hauptmerkmale

• 2–4× schnellere Inferenz im Vergleich zu OpenAI Whisper
• Geringerer Speicherbedarf dank Unterstützung für Quantisierung
• Gleiche Genauigkeit wie die Original-Whisper-Modelle
• GPU- und CPU-Unterstützung mit optimierten Backends
• Stapelverarbeitung für mehrere Dateien
• Wortgenaue Zeitstempel
• Quantisierungsoptionen (FP32, FP16, INT8, INT8_FLOAT16)
• Sprachaktivitätserkennung (VAD) zur Filterung

Funktionsweise

• Schnellere Matrixoperationen mit optimierten BLAS-Bibliotheken
• Besseres Speichermanagement mit geringerem Overhead
• Quantisierung für geringeren Speicherverbrauch
• Stapelverarbeitung zur Durchsatzoptimierung

---

Faster-Whisper vs. OpenAI Whisper

Leistungsvergleich

Merkmal	OpenAI Whisper	Faster-Whisper
Geschwindigkeit	Referenz	2–4× schneller
Speichernutzung	Höher	Geringer (mit Quantisierung)
Genauigkeit	Hoch	Gleich (identische Modelle)
GPU-Unterstützung	Ja	Ja (optimiert)
CPU-Unterstützung	Ja	Ja (optimiert)
Quantisierung	Begrenzt	Vollständig (INT8, FP16)
Stapelverarbeitung	Manuell	Integriert
Installation	Einfach	Einfach (inkl. CTranslate2)

Wann Faster-Whisper nutzen

• Sie schnellere Transkription für Produktionslast benötigen
• Sie mehrere Dateien im Batch verarbeiten
• Sie ressourcenbeschränkte Systeme einsetzen (INT8 nutzen)
• Sie Echtzeit- oder nahezu Echtzeit-Anwendungen bauen
• Sie weniger Arbeitsspeicher für den Einsatz brauchen

• Sie maximale Kompatibilität mit bestehendem Code brauchen
• Sie feinabgestimmte Modelle nutzen (faster-whisper erfordert Konvertierung)
• Sie eine einfachere API bevorzugen (faster-whisper ist jedoch ähnlich)
• Sie experimentelle Funktionen nutzen, die zuerst in OpenAI Whisper erscheinen

---

Installation

Voraussetzungen

• Python 3.9+ (erforderlich)
• FFmpeg (optional – faster-whisper nutzt PyAV, für manche Formate kann FFmpeg nötig sein)
• NVIDIA-GPU (optional, für GPU-Beschleunigung)

Basisinstallation

pip install faster-whisper

• Paket faster-whisper
• ctranslate2 (CTranslate2-Inferenz-Engine)
• pyav (Audiodekodierung, ersetzt die FFmpeg-Abhängigkeit)

GPU-Installation (NVIDIA CUDA)

pip install nvidia-cublas-cu12 nvidia-cudnn-cu12==9.*

export LD_LIBRARY_PATH=$(python3 -c 'import os; import nvidia.cublas.lib; import nvidia.cudnn.lib; print(os.path.dirname(nvidia.cublas.lib.__file__) + ":" + os.path.dirname(nvidia.cudnn.lib.__file__))')

pip install ctranslate2==3.24.0 faster-whisper

Installation prüfen

from faster_whisper import WhisperModel

# Test basic import
print("Faster-whisper installed successfully!")

---

Grundlegende Nutzung

Einfache Transkription

from faster_whisper import WhisperModel

# Load model (automatically downloads if not present)
model = WhisperModel("base", device="cpu", compute_type="int8")

# Transcribe audio
segments, info = model.transcribe("audio.mp3")

# Print detected language
print(f"Detected language: {info.language} (probability: {info.language_probability:.2f})")

# Print transcription
for segment in segments:
    print(f"[{segment.start:.2f}s -> {segment.end:.2f}s] {segment.text}")

Vollständigen Text erhalten

from faster_whisper import WhisperModel

model = WhisperModel("base")
segments, info = model.transcribe("audio.mp3")

# Collect all text
full_text = " ".join([segment.text for segment in segments])
print(full_text)

Mit Wort-Zeitstempeln

from faster_whisper import WhisperModel

model = WhisperModel("base", device="cpu", compute_type="int8")

segments, info = model.transcribe(
    "audio.mp3",
    word_timestamps=True,
    beam_size=5
)

for segment in segments:
    print(f"[{segment.start:.2f}s - {segment.end:.2f}s] {segment.text}")
    
    # Word-level timestamps
    for word in segment.words:
        print(f"  {word.word} [{word.start:.2f}s - {word.end:.2f}s]")

---

Gerät- und Compute-Typ-Optionen

Geräteoptionen

• device="cpu" – CPU-Inferenz (funktioniert überall)
• device="cuda" – GPU-Inferenz (NVIDIA-GPU und CUDA erforderlich)

Compute-Typen

Compute-Typ	Geschwindigkeit	Speicher	Genauigkeit	Anwendungsfall
int8	Am schnellsten	Am geringsten	Etwas geringer	CPU, ressourcenbeschränkt
int8_float16	Sehr schnell	Gering	Hoch	GPU mit begrenztem VRAM
float16	Schnell	Mittel	Hoch	GPU (empfohlen)
float32	Am langsamsten	Am höchsten	Am höchsten	Maximale Genauigkeit

Beispiele nach Hardware

# Best for CPU: INT8
model = WhisperModel("base", device="cpu", compute_type="int8")

# Best for GPU: FP16
model = WhisperModel("large-v2", device="cuda", compute_type="float16")

# Use INT8_FLOAT16 for large models
model = WhisperModel("large-v2", device="cuda", compute_type="int8_float16")

# Use FP32 (slower but most accurate)
model = WhisperModel("large-v2", device="cuda", compute_type="float32")

---

Erweiterte Funktionen

1. Stapelverarbeitung

from faster_whisper import WhisperModel
from pathlib import Path

model = WhisperModel("base", device="cuda", compute_type="float16")

audio_files = ["audio1.mp3", "audio2.mp3", "audio3.mp3"]

for audio_file in audio_files:
    print(f"Transcribing: {audio_file}")
    segments, info = model.transcribe(audio_file)
    
    text = " ".join([seg.text for seg in segments])
    print(f"Result: {text[:100]}...")
    print()

2. Sprachaktivitätserkennung (VAD)

from faster_whisper import WhisperModel

model = WhisperModel("base")

segments, info = model.transcribe(
    "audio.mp3",
    vad_filter=True,  # Enable VAD filtering
    vad_parameters=dict(
        min_silence_duration_ms=500,  # Minimum silence duration
        threshold=0.5  # VAD threshold
    )
)

for segment in segments:
    print(f"[{segment.start:.2f}s] {segment.text}")

3. Sprachangabe

from faster_whisper import WhisperModel

model = WhisperModel("base")

# Specify language (faster and more accurate)
segments, info = model.transcribe(
    "audio.mp3",
    language="en"  # English
)

# Or let it auto-detect
segments, info = model.transcribe("audio.mp3")  # Auto-detect
print(f"Detected: {info.language}")

4. Beam-Größe und weitere Parameter

from faster_whisper import WhisperModel

model = WhisperModel("base")

segments, info = model.transcribe(
    "audio.mp3",
    beam_size=5,  # Higher = more accurate but slower (default: 5)
    best_of=5,    # Number of candidates to consider
    temperature=0.0,  # Lower = more deterministic
    condition_on_previous_text=True,  # Use context from previous segments
    initial_prompt="This is a technical meeting about AI and machine learning."
)

5. Eigene Modellpfade

from faster_whisper import WhisperModel

# Use local model directory
model = WhisperModel(
    "base",
    device="cpu",
    compute_type="int8",
    download_root="./models"  # Custom download directory
)

# Or specify full path to converted model
model = WhisperModel(
    "/path/to/converted/model",
    device="cuda",
    compute_type="float16"
)

---

Leistungs-Benchmarks

GPU-Leistung (NVIDIA RTX 3070 Ti)

Konfiguration	Zeit	VRAM-Nutzung	Speedup
OpenAI Whisper (FP16, beam=5)	~2m 23s	~4708 MB	Referenz
Faster-whisper (FP16, beam=5)	~1m 03s	~4525 MB	2,3× schneller
Faster-whisper (INT8, beam=5)	~59s	~2926 MB	2,4× schneller
Faster-whisper (FP16, batch=8)	~17s	~6090 MB	8,4× schneller
Faster-whisper (INT8, batch=8)	~16s	~4500 MB	8,9× schneller

CPU-Leistung (Intel Core i7-12700K)

Konfiguration	Zeit	RAM-Nutzung	Speedup
OpenAI Whisper (FP32, beam=5)	~6m 58s	~2335 MB	Referenz
Faster-whisper (FP32, beam=5)	~2m 37s	~2257 MB	2,7× schneller
Faster-whisper (INT8, beam=5)	~1m 42s	~1477 MB	4,1× schneller
Faster-whisper (FP32, batch=8)	~1m 06s	~4230 MB	6,3× schneller
Faster-whisper (INT8, batch=8)	~51s	~3608 MB	8,2× schneller

Wichtige Erkenntnisse

• Stapelverarbeitung liefert den größten Speedup (8×+ auf der GPU)
• INT8-Quantisierung reduziert den Speicher um ~40 % bei minimaler Genauigkeitseinbuße
• GPU-Beschleunigung ist für große Modelle und Stapelverarbeitung wesentlich
• CPU mit INT8 ist für kleinere Modelle und Einzeldateien praktikabel

---

Vollständiges Beispiel: produktionsreife Transkription

from faster_whisper import WhisperModel
from pathlib import Path
import json
from datetime import datetime

class TranscriptionService:
    """Production-ready transcription service using faster-whisper."""
    
    def __init__(self, model_size="base", device="cpu", compute_type="int8"):
        """Initialize the transcription service."""
        print(f"Loading model: {model_size} on {device} ({compute_type})")
        self.model = WhisperModel(
            model_size,
            device=device,
            compute_type=compute_type
        )
        print("Model loaded successfully!")
    
    def transcribe_file(self, audio_path, output_format="txt", **kwargs):
        """
        Transcribe an audio file.
        
        Args:
            audio_path: Path to audio file
            output_format: Output format (txt, json, srt, vtt)
            **kwargs: Additional transcription parameters
        """
        audio_path = Path(audio_path)
        if not audio_path.exists():
            raise FileNotFoundError(f"Audio file not found: {audio_path}")
        
        print(f"Transcribing: {audio_path.name}")
        
        # Transcribe
        segments, info = self.model.transcribe(
            str(audio_path),
            word_timestamps=True,
            **kwargs
        )
        
        # Collect results
        result = {
            "file": str(audio_path),
            "language": info.language,
            "language_probability": info.language_probability,
            "duration": info.duration,
            "segments": []
        }
        
        full_text_parts = []
        for segment in segments:
            segment_data = {
                "start": segment.start,
                "end": segment.end,
                "text": segment.text,
                "words": [
                    {
                        "word": word.word,
                        "start": word.start,
                        "end": word.end,
                        "probability": word.probability
                    }
                    for word in segment.words
                ]
            }
            result["segments"].append(segment_data)
            full_text_parts.append(segment.text)
        
        result["text"] = " ".join(full_text_parts)
        
        # Save based on format
        output_path = audio_path.parent / f"{audio_path.stem}_transcript"
        
        if output_format == "txt":
            self._save_txt(result, output_path.with_suffix(".txt"))
        elif output_format == "json":
            self._save_json(result, output_path.with_suffix(".json"))
        elif output_format == "srt":
            self._save_srt(result, output_path.with_suffix(".srt"))
        elif output_format == "vtt":
            self._save_vtt(result, output_path.with_suffix(".vtt"))
        
        print(f"✓ Transcription saved: {output_path}.{output_format}")
        return result
    
    def _save_txt(self, result, path):
        """Save as plain text."""
        with open(path, "w", encoding="utf-8") as f:
            f.write(result["text"])
    
    def _save_json(self, result, path):
        """Save as JSON."""
        with open(path, "w", encoding="utf-8") as f:
            json.dump(result, f, indent=2, ensure_ascii=False)
    
    def _save_srt(self, result, path):
        """Save as SRT subtitles."""
        with open(path, "w", encoding="utf-8") as f:
            for i, seg in enumerate(result["segments"], start=1):
                start = self._format_srt_time(seg["start"])
                end = self._format_srt_time(seg["end"])
                f.write(f"{i}\n{start} --> {end}\n{seg['text']}\n\n")
    
    def _save_vtt(self, result, path):
        """Save as WebVTT."""
        with open(path, "w", encoding="utf-8") as f:
            f.write("WEBVTT\n\n")
            for seg in result["segments"]:
                start = self._format_vtt_time(seg["start"])
                end = self._format_vtt_time(seg["end"])
                f.write(f"{start} --> {end}\n{seg['text']}\n\n")
    
    def _format_srt_time(self, seconds):
        """Format time for SRT."""
        hours = int(seconds // 3600)
        minutes = int((seconds % 3600) // 60)
        secs = int(seconds % 60)
        millis = int((seconds % 1) * 1000)
        return f"{hours:02d}:{minutes:02d}:{secs:02d},{millis:03d}"
    
    def _format_vtt_time(self, seconds):
        """Format time for VTT."""
        hours = int(seconds // 3600)
        minutes = int((seconds % 3600) // 60)
        secs = int(seconds % 60)
        millis = int((seconds % 1) * 1000)
        return f"{hours:02d}:{minutes:02d}:{secs:02d}.{millis:03d}"

# Usage
if __name__ == "__main__":
    # Initialize service
    service = TranscriptionService(
        model_size="base",
        device="cpu",  # Change to "cuda" for GPU
        compute_type="int8"  # Use "float16" for GPU
    )
    
    # Transcribe file
    result = service.transcribe_file(
        "meeting.mp3",
        output_format="json",
        beam_size=5,
        language="en"
    )
    
    print(f"\nLanguage: {result['language']}")
    print(f"Duration: {result['duration']:.2f}s")
    print(f"Text: {result['text'][:200]}...")

---

Bewährte Praktiken

1. Passende Modellgröße wählen

# For speed (CPU)
model = WhisperModel("tiny", device="cpu", compute_type="int8")

# For balance
model = WhisperModel("base", device="cpu", compute_type="int8")

# For accuracy (GPU recommended)
model = WhisperModel("large-v2", device="cuda", compute_type="float16")

2. Hardware optimieren

model = WhisperModel("base", device="cpu", compute_type="int8")

model = WhisperModel("large-v2", device="cuda", compute_type="float16")

model = WhisperModel("medium", device="cuda", compute_type="int8_float16")

3. Stapelverarbeitung für mehrere Dateien

# Process multiple files efficiently
audio_files = ["file1.mp3", "file2.mp3", "file3.mp3"]
model = WhisperModel("base", device="cuda", compute_type="float16")

for audio_file in audio_files:
    segments, info = model.transcribe(audio_file)
    # Process results...

4. VAD bei verrauschtem Audio aktivieren

segments, info = model.transcribe(
    "noisy_audio.mp3",
    vad_filter=True,
    vad_parameters=dict(
        min_silence_duration_ms=1000,
        threshold=0.5
    )
)

5. Sprache angeben, wenn bekannt

# Faster and more accurate when language is known
segments, info = model.transcribe(
    "audio.mp3",
    language="en"  # Specify instead of auto-detect
)

6. Modellinstanzen wiederverwenden

# Load model once, reuse for multiple files
model = WhisperModel("base")

# Process multiple files with same model
for audio_file in audio_files:
    segments, info = model.transcribe(audio_file)

---

Migration von OpenAI Whisper

Code-Vergleich

import whisper

model = whisper.load_model("base")
result = model.transcribe("audio.mp3")
print(result["text"])

from faster_whisper import WhisperModel

model = WhisperModel("base", device="cpu", compute_type="int8")
segments, info = model.transcribe("audio.mp3")
text = " ".join([seg.text for seg in segments])
print(text)

Wichtige Unterschiede

1. Modell laden: WhisperModel() statt whisper.load_model()
2. Rückgabeformat: Tupel (segments, info) statt Dictionary
3. Segmente: Iterator von Segmentobjekten statt Liste
4. Device/Compute-Typ: device und compute_type müssen angegeben werden
5. Textzugriff: Segmente müssen für den Volltext verkettet werden

Hilfsfunktion für die Migration

def convert_to_whisper_format(segments, info):
    """Convert faster-whisper output to OpenAI Whisper format."""
    return {
        "text": " ".join([seg.text for seg in segments]),
        "language": info.language,
        "segments": [
            {
                "id": i,
                "start": seg.start,
                "end": seg.end,
                "text": seg.text,
                "words": [
                    {
                        "word": word.word,
                        "start": word.start,
                        "end": word.end
                    }
                    for word in seg.words
                ] if hasattr(seg, 'words') else []
            }
            for i, seg in enumerate(segments)
        ]
    }

# Usage
segments, info = model.transcribe("audio.mp3", word_timestamps=True)
result = convert_to_whisper_format(segments, info)
# Now compatible with OpenAI Whisper format

---

Fehlerbehebung

Problem 1: CUDA – nicht genug Speicher

# Use smaller model
model = WhisperModel("base", device="cuda", compute_type="float16")

# Or use INT8 quantization
model = WhisperModel("large-v2", device="cuda", compute_type="int8_float16")

# Or use CPU
model = WhisperModel("large-v2", device="cpu", compute_type="int8")

Problem 2: Langsame CPU-Leistung

# Use INT8 quantization
model = WhisperModel("base", device="cpu", compute_type="int8")

# Use smaller model
model = WhisperModel("tiny", device="cpu", compute_type="int8")

# Reduce beam size
segments, info = model.transcribe("audio.mp3", beam_size=1)

Problem 3: CUDA-Bibliotheken nicht gefunden

# Install CUDA libraries
pip install nvidia-cublas-cu12 nvidia-cudnn-cu12==9.*

# Set library path
export LD_LIBRARY_PATH=$(python3 -c 'import os; import nvidia.cublas.lib; import nvidia.cudnn.lib; print(os.path.dirname(nvidia.cublas.lib.__file__) + ":" + os.path.dirname(nvidia.cudnn.lib.__file__))')

Problem 4: Modell-Download schlägt fehl

# Specify download directory
model = WhisperModel(
    "base",
    download_root="./models",  # Custom directory
    local_files_only=False
)

# Or download manually from Hugging Face
# Then use local path
model = WhisperModel("/path/to/local/model")

---

Wann Faster-Whisper einsetzen

Faster-whisper nutzen, wenn:

OpenAI Whisper nutzen, wenn:

---

Fazit

• INT8 für CPU und ressourcenbeschränkte Systeme
• FP16 für GPU mit ausreichend VRAM
• Stapelverarbeitung für mehrere Dateien aktivieren
• Sprache angeben, wenn bekannt, für bessere Leistung
• Modellinstanzen wiederverwenden für mehrere Transkriptionen

---