Tutorial OpenAI Whisper: guida completa alla trascrizione speech-to-text

---

Cos'è OpenAI Whisper?

Caratteristiche principali

• Supporto multilingue — oltre 99 lingue
• Trascrizione speech-to-text — converte l'audio in testo
• Traduzione della voce — traduce la voce direttamente in inglese
• Rilevamento della lingua — rileva automaticamente la lingua parlata
• Generazione di timestamp — timestamp a livello di parola e di segmento
• Open source e gratuito — licenza MIT, nessun costo API
• Utilizzabile offline — viene eseguito in locale sulla tua macchina
• Formati multipli — supporta vari formati audio/video

---

Dimensioni dei modelli Whisper spiegate

Modello	Parametri	Velocità	Accuratezza	Memoria	Caso d'uso
tiny	39M	⭐⭐⭐⭐⭐	⭐⭐	~1 GB	Test rapidi, demo
base	74M	⭐⭐⭐⭐	⭐⭐⭐	~1 GB	Audio semplice, attività veloci
small	244M	⭐⭐⭐	⭐⭐⭐⭐	~2 GB	Uso generale, bilanciato
medium	769M	⭐⭐	⭐⭐⭐⭐⭐	~5 GB	Audio rumoroso, alta accuratezza
large	1550M	⭐	⭐⭐⭐⭐⭐⭐	~10 GB	Massima accuratezza, produzione

• Per la velocità: usa tiny o base
• Per un compromesso: usa small o medium
• Per l'accuratezza: usa large o large-v3
• Per la produzione: spesso medium o large-v2

---

Prerequisiti

• Python 3.8 o successivo (consigliato Python 3.9+)
• pip come gestore pacchetti
• FFmpeg installato (per l'elaborazione audio/video)
• (Opzionale) GPU NVIDIA con CUDA per elaborazione più veloce
• (Opzionale) 4 GB+ di RAM per il modello base, 10 GB+ per il modello large

---

Passaggio 1: Installazione

Installare Whisper

pip install openai-whisper

pip install openai-whisper==20231117

Installare FFmpeg

brew install ffmpeg

sudo apt update
sudo apt install ffmpeg

1. Scarica FFmpeg da ffmpeg.org
2. Estrai e aggiungi al PATH di sistema
3. Oppure usa: choco install ffmpeg (con Chocolatey)

ffmpeg -version
whisper --version

---

Passaggio 2: Utilizzo di base — Python

Trascrizione semplice

import whisper

# Load model (downloads automatically on first use)
model = whisper.load_model("base")

# Transcribe audio file
result = model.transcribe("audio.mp3")

# Print transcription
print(result["text"])

Hello everyone, welcome to today's meeting. We will discuss the project timeline and upcoming milestones.

Esempio completo con gestione degli errori

import whisper
import os

def transcribe_audio(audio_path, model_size="base"):
    """
    Transcribe an audio file using Whisper.
    
    Args:
        audio_path (str): Path to the audio file
        model_size (str): Whisper model size (tiny, base, small, medium, large)
    
    Returns:
        dict: Transcription result with text and segments
    """
    try:
        # Check if audio file exists
        if not os.path.exists(audio_path):
            raise FileNotFoundError(f"Audio file not found: {audio_path}")
        
        # Load the Whisper model
        print(f"Loading Whisper model: {model_size}")
        model = whisper.load_model(model_size)
        
        # Transcribe the audio
        print(f"Transcribing: {audio_path}")
        result = model.transcribe(audio_path)
        
        print(f"✓ Transcription complete!")
        print(f"  Language: {result['language']}")
        print(f"  Duration: {result['segments'][-1]['end']:.2f}s")
        
        return result
    
    except Exception as e:
        print(f"Error during transcription: {str(e)}")
        return None

# Example usage
if __name__ == "__main__":
    audio_file = "meeting.mp3"
    result = transcribe_audio(audio_file, model_size="base")
    
    if result:
        print("\n" + "="*50)
        print("TRANSCRIPTION:")
        print("="*50)
        print(result["text"])

---

Passaggio 3: Rilevamento e specifica della lingua

Rilevamento automatico della lingua

import whisper

model = whisper.load_model("base")
result = model.transcribe("audio.mp3")

print(f"Detected language: {result['language']}")
print(f"Language probability: {result.get('language_probability', 0):.2%}")
print(f"\nTranscription:\n{result['text']}")

Specificare la lingua (più veloce e più accurato)

import whisper

model = whisper.load_model("base")

# Specify language
result_en = model.transcribe("audio.mp3", language="en")  # English
result_zh = model.transcribe("audio.mp3", language="zh")   # Chinese
result_es = model.transcribe("audio.mp3", language="es")  # Spanish
result_fr = model.transcribe("audio.mp3", language="fr")  # French
result_de = model.transcribe("audio.mp3", language="de")  # German
result_ja = model.transcribe("audio.mp3", language="ja")   # Japanese

print(result_en["text"])

• en - English
• zh - Chinese
• es - Spanish
• fr - French
• de - German
• ja - Japanese
• ko - Korean
• pt - Portuguese
• ru - Russian
• it - Italian

---

Passaggio 4: Timestamp e segmenti

Accedere ai segmenti con timestamp

import whisper

model = whisper.load_model("base")
result = model.transcribe("audio.mp3")

# Print full transcription
print("Full Text:")
print(result["text"])

# Print segments with timestamps
print("\n" + "="*50)
print("Segments with Timestamps:")
print("="*50)

for segment in result["segments"]:
    start = segment["start"]
    end = segment["end"]
    text = segment["text"].strip()
    print(f"[{start:6.2f}s - {end:6.2f}s] {text}")

Full Text:
Hello everyone, welcome to today's meeting. We will discuss the project timeline.

==================================================
Segments with Timestamps:
==================================================
[  0.00s -   5.20s] Hello everyone, welcome to today's meeting.
[  5.20s -  12.50s] We will discuss the project timeline.

Formattare i timestamp come timecode

def format_timestamp(seconds):
    """Format seconds to HH:MM:SS."""
    hours = int(seconds // 3600)
    minutes = int((seconds % 3600) // 60)
    secs = int(seconds % 60)
    return f"{hours:02d}:{minutes:02d}:{secs:02d}"

for segment in result["segments"]:
    start_time = format_timestamp(segment["start"])
    end_time = format_timestamp(segment["end"])
    print(f"[{start_time} - {end_time}] {segment['text']}")

Timestamp a livello di parola

import whisper

model = whisper.load_model("base")

result = model.transcribe(
    "audio.mp3",
    word_timestamps=True  # Enable word-level timestamps
)

for segment in result["segments"]:
    print(f"\n[{segment['start']:.2f}s - {segment['end']:.2f}s]")
    print(f"Text: {segment['text']}")
    
    # Word-level timestamps
    if "words" in segment:
        print("Words:")
        for word in segment["words"]:
            print(f"  {word['word']} [{word['start']:.2f}s - {word['end']:.2f}s]")

---

Passaggio 5: Traduzione della voce

import whisper

model = whisper.load_model("base")

# Translate to English (regardless of source language)
result = model.transcribe("spanish_audio.mp3", task="translate")

print("Translated to English:")
print(result["text"])

# Original transcription (in original language)
result_original = model.transcribe("spanish_audio.mp3", task="transcribe")
print("\nOriginal language transcription:")
print(result_original["text"])

• Riunioni internazionali
• Elaborazione di contenuti multilingue
• Localizzazione dei contenuti
• Materiali per l'apprendimento linguistico

---

Passaggio 6: Parametri avanzati

Temperatura e beam size

import whisper

model = whisper.load_model("base")

result = model.transcribe(
    "audio.mp3",
    temperature=0.0,        # Lower = more deterministic (0.0 recommended)
    beam_size=5,            # Higher = more accurate but slower (default: 5)
    best_of=5,              # Number of candidates to consider
    patience=1.0,           # Beam search patience
    condition_on_previous_text=True,  # Use context from previous segments
    initial_prompt="This is a technical meeting about AI and machine learning."  # Context prompt
)

Valori di temperatura

• temperature=0.0 — Più deterministico, consigliato
• temperature=0.2-0.4 — Leggermente più variazione
• temperature=1.0 — Più creativo, meno accurato

Prompt iniziale per il contesto

result = model.transcribe(
    "technical_meeting.mp3",
    initial_prompt="This meeting discusses API endpoints, microservices, Kubernetes, and CI/CD pipelines."
)

result = model.transcribe(
    "medical_audio.mp3",
    initial_prompt="This is a medical consultation discussing patient symptoms and treatment options."
)

---

Passaggio 7: Interfaccia a riga di comando (CLI)

Uso CLI di base

whisper audio.mp3

Specificare il modello

whisper audio.mp3 --model small
whisper audio.mp3 --model medium
whisper audio.mp3 --model large-v2

Specificare la lingua

whisper audio.mp3 --language en
whisper audio.mp3 --language zh

Formati di output

# SRT subtitles
whisper audio.mp3 --output_format srt

# VTT subtitles
whisper audio.mp3 --output_format vtt

# Text file
whisper audio.mp3 --output_format txt

# JSON (with all metadata)
whisper audio.mp3 --output_format json

# TSV (tab-separated values)
whisper audio.mp3 --output_format tsv

Opzioni avanzate della CLI

# Full example with all options
whisper audio.mp3 \
  --model medium \
  --language en \
  --task transcribe \
  --output_format srt \
  --output_dir ./transcripts \
  --verbose True \
  --temperature 0.0 \
  --beam_size 5 \
  --best_of 5 \
  --fp16 True

Riferimento opzioni CLI

Opzione	Descrizione	Predefinito
--model	Dimensione del modello (tiny, base, small, medium, large)	base
--language	Codice lingua (en, zh, es, ecc.)	Rilevamento automatico
--task	transcribe o translate	transcribe
--output_format	Formato di output (txt, srt, vtt, json, tsv)	txt
--output_dir	Cartella di output	Cartella corrente
--temperature	Temperatura per il campionamento	0.0
--beam_size	Beam size per la beam search	5
--best_of	Numero di candidati	5
--fp16	Usa precisione FP16 (GPU)	True
--verbose	Output dettagliato	False

---

Passaggio 8: Formati audio e video supportati

Formati supportati

• Audio: MP3, WAV, M4A, FLAC, OGG, AAC, WMA
• Video: MP4, AVI, MKV, MOV, WebM, FLV
• Streaming: può elaborare flussi audio

Esempi di formato

import whisper

model = whisper.load_model("base")

# Audio formats
model.transcribe("audio.mp3")
model.transcribe("audio.wav")
model.transcribe("audio.m4a")
model.transcribe("audio.flac")

# Video formats (extracts audio automatically)
model.transcribe("video.mp4")
model.transcribe("video.mkv")
model.transcribe("video.webm")

---

Passaggio 9: Esempio completo per la produzione

import whisper
import json
from pathlib import Path
from datetime import datetime

class WhisperTranscriber:
    """Production-ready Whisper transcription service."""
    
    def __init__(self, model_size="base"):
        """Initialize transcriber with specified model."""
        print(f"Loading Whisper model: {model_size}")
        self.model = whisper.load_model(model_size)
        print("✓ Model loaded successfully")
    
    def transcribe_file(self, audio_path, output_dir="transcripts", **kwargs):
        """
        Transcribe audio file and save results.
        
        Args:
            audio_path: Path to audio file
            output_dir: Directory to save outputs
            **kwargs: Additional transcribe parameters
        """
        audio_path = Path(audio_path)
        if not audio_path.exists():
            raise FileNotFoundError(f"Audio file not found: {audio_path}")
        
        output_path = Path(output_dir)
        output_path.mkdir(exist_ok=True)
        
        print(f"\nTranscribing: {audio_path.name}")
        
        # Transcribe
        result = self.model.transcribe(
            str(audio_path),
            word_timestamps=True,
            **kwargs
        )
        
        # Prepare output data
        output_data = {
            "file": str(audio_path),
            "transcribed_at": datetime.now().isoformat(),
            "language": result["language"],
            "language_probability": result.get("language_probability", 0),
            "duration": result["segments"][-1]["end"] if result["segments"] else 0,
            "text": result["text"],
            "segments": result["segments"]
        }
        
        # Save outputs
        base_name = audio_path.stem
        
        # Save as text
        text_file = output_path / f"{base_name}.txt"
        with open(text_file, "w", encoding="utf-8") as f:
            f.write(result["text"])
        
        # Save as JSON
        json_file = output_path / f"{base_name}.json"
        with open(json_file, "w", encoding="utf-8") as f:
            json.dump(output_data, f, indent=2, ensure_ascii=False)
        
        # Save as SRT
        srt_file = output_path / f"{base_name}.srt"
        self._save_srt(result["segments"], srt_file)
        
        print(f"✓ Transcription saved:")
        print(f"  - Text: {text_file}")
        print(f"  - JSON: {json_file}")
        print(f"  - SRT: {srt_file}")
        
        return output_data
    
    def _save_srt(self, segments, output_path):
        """Save segments as SRT subtitle file."""
        with open(output_path, "w", encoding="utf-8") as f:
            for i, segment in enumerate(segments, start=1):
                start = self._format_srt_time(segment["start"])
                end = self._format_srt_time(segment["end"])
                text = segment["text"].strip()
                f.write(f"{i}\n{start} --> {end}\n{text}\n\n")
    
    def _format_srt_time(self, seconds):
        """Format seconds to SRT timestamp."""
        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__":
    transcriber = WhisperTranscriber(model_size="base")
    
    result = transcriber.transcribe_file(
        "meeting.mp3",
        output_dir="transcripts",
        language="en",
        temperature=0.0
    )
    
    print(f"\nLanguage: {result['language']}")
    print(f"Duration: {result['duration']:.2f}s")
    print(f"\nTranscription preview:")
    print(result['text'][:200] + "...")

---

Passaggio 10: Best practice

1. Scegliere il modello giusto

# For speed (testing, demos)
model = whisper.load_model("tiny")

# For balance (general use)
model = whisper.load_model("base")  # or "small"

# For accuracy (production)
model = whisper.load_model("medium")  # or "large-v2"

2. Specificare la lingua quando è nota

# Faster and more accurate
result = model.transcribe("audio.mp3", language="en")

# Instead of auto-detection
result = model.transcribe("audio.mp3")  # Slower

3. Usare la temperatura appropriata

# Recommended for most cases
result = model.transcribe("audio.mp3", temperature=0.0)

# For creative content (not recommended for transcription)
result = model.transcribe("audio.mp3", temperature=0.2)

4. Fornire contesto con un prompt iniziale

# Technical content
result = model.transcribe(
    "meeting.mp3",
    initial_prompt="This meeting discusses software architecture, APIs, and deployment strategies."
)

# Medical content
result = model.transcribe(
    "consultation.mp3",
    initial_prompt="This is a medical consultation about patient symptoms and treatment."
)

5. Riutilizzare le istanze del modello

# Load once, reuse multiple times
model = whisper.load_model("base")

# Process multiple files
for audio_file in ["file1.mp3", "file2.mp3", "file3.mp3"]:
    result = model.transcribe(audio_file)
    # Process result...

6. Gestire file audio molto lunghi

import whisper
from pydub import AudioSegment

def transcribe_long_audio(audio_path, chunk_length_ms=600000):  # 10 minutes
    """Transcribe long audio by splitting into chunks."""
    model = whisper.load_model("base")
    
    # Load audio
    audio = AudioSegment.from_file(audio_path)
    duration_ms = len(audio)
    
    all_text = []
    all_segments = []
    
    # Process in chunks
    for i in range(0, duration_ms, chunk_length_ms):
        chunk = audio[i:i + chunk_length_ms]
        chunk_path = f"chunk_{i}.wav"
        chunk.export(chunk_path, format="wav")
        
        result = model.transcribe(chunk_path)
        all_text.append(result["text"])
        all_segments.extend(result["segments"])
        
        # Clean up chunk file
        os.remove(chunk_path)
    
    return {
        "text": " ".join(all_text),
        "segments": all_segments
    }

---

Problemi comuni e soluzioni

Problema 1: FFmpeg non trovato

# Install FFmpeg
# macOS
brew install ffmpeg

# Ubuntu/Debian
sudo apt install ffmpeg

# Verify
ffmpeg -version

Problema 2: Memoria insufficiente

# Use smaller model
model = whisper.load_model("base")  # Instead of "large"

# Or use CPU
import torch
model = whisper.load_model("base", device="cpu")

# Or process in chunks (see above)

Problema 3: Trascrizione lenta

# Use GPU if available
import torch
device = "cuda" if torch.cuda.is_available() else "cpu"
model = whisper.load_model("base", device=device)

# Use smaller model
model = whisper.load_model("tiny")  # or "base"

# Reduce beam size (faster but slightly less accurate)
result = model.transcribe("audio.mp3", beam_size=1)

Problema 4: Scarsa accuratezza

# Use larger model
model = whisper.load_model("medium")  # or "large"

# Specify language
result = model.transcribe("audio.mp3", language="en")

# Provide context
result = model.transcribe(
    "audio.mp3",
    initial_prompt="Context about the audio content..."
)

# Use optimal settings
result = model.transcribe(
    "audio.mp3",
    temperature=0.0,
    beam_size=5,
    best_of=5
)

---

Casi d'uso

1. Trascrizione di podcast

model = whisper.load_model("medium")
result = model.transcribe("podcast.mp3", language="en")

# Save transcript
with open("podcast_transcript.txt", "w") as f:
    f.write(result["text"])

2. Generazione sottotitoli per YouTube

model = whisper.load_model("base")
result = model.transcribe("video.mp4", language="en")

# Generate SRT
# (Use CLI: whisper video.mp4 --output_format srt)

3. Appunti da riunione

model = whisper.load_model("base")
result = model.transcribe(
    "meeting.mp3",
    language="en",
    initial_prompt="This is a business meeting discussing project updates and deadlines."
)

# Save with timestamps
for segment in result["segments"]:
    print(f"[{segment['start']:.0f}s] {segment['text']}")

4. Trascrizione di interviste

model = whisper.load_model("medium")
result = model.transcribe("interview.mp3", language="en")

# Export for editing
with open("interview.txt", "w") as f:
    for segment in result["segments"]:
        f.write(f"[{segment['start']:.2f}s] {segment['text']}\n")

5. Traduzione di contenuti multilingue

model = whisper.load_model("base")

# Translate to English
result = model.transcribe("spanish_audio.mp3", task="translate")
print(result["text"])  # English translation

---

Whisper rispetto alle alternative

Funzionalità	Whisper	API cloud	Faster-Whisper
Costo	Gratis	A pagamento al minuto	Gratis
Offline	✅	❌	✅
Velocità	Media	Veloce	Veloce (2-4×)
Accuratezza	Alta	Alta	Alta (stessa)
Configurazione	Facile	Molto facile	Facile
In tempo reale	❌	✅	❌
Privacy	✅ In locale	❌ Cloud	✅ In locale

• Vuoi una trascrizione gratuita e offline
• La privacy è importante
• Hai controllo dell'infrastruttura
• Elabori file in batch o contenuti archiviati

• Ti serve la trascrizione in tempo reale
• Vuoi un'infrastruttura gestita
• Hai un budget per i costi API
• Ti serve il supporto enterprise

---

Passi successivi

• Whisper Python Example - Esempi Python più dettagliati
• Faster-Whisper Guide - Trascrizione 2-4× più veloce
• Whisper Accuracy Tips - Migliora la qualità della trascrizione
• Whisper Transcript Formatting - Formatta gli output (SRT, VTT, JSON)
• Whisper for Meetings - Trascrizione specifica per riunioni

---

Conclusione

• Whisper supporta oltre 99 lingue con alta accuratezza
• Scegli la dimensione del modello adeguata alle tue esigenze
• Specifica la lingua quando è nota per prestazioni migliori
• Usa i timestamp a livello di parola per un timing preciso
• Riutilizza le istanze del modello per più file
• Valuta faster-whisper per distribuzioni in produzione

---