Whisper Transcription

Whisper Transcription

OpenAI Whisper provides accurate speech-to-text with word-level timestamps.

Installation

pip install openai-whisper

Model Selection

Use the tiny model for fast transcription - it's sufficient for most tasks and runs much faster:

| Model | Size | Speed | Accuracy | |-------|------|-------|----------| | tiny | 39 MB | Fastest | Good for clear speech | | base | 74 MB | Fast | Better accuracy | | small | 244 MB | Medium | High accuracy |

Recommendation: Start with tiny - it handles clear interview/podcast audio well.

Basic Usage with Word Timestamps

import whisper
import json

def transcribe_with_timestamps(audio_path, output_path):
    """
    Transcribe audio and get word-level timestamps.

    Args:
        audio_path: Path to audio/video file
        output_path: Path to save JSON output
    """
    # Use tiny model for speed
    model = whisper.load_model("tiny")

    # Transcribe with word timestamps
    result = model.transcribe(
        audio_path,
        word_timestamps=True,
        language="en"  # Specify language for better accuracy
    )

    # Extract words with timestamps
    words = []
    for segment in result["segments"]:
        if "words" in segment:
            for word_info in segment["words"]:
                words.append({
                    "word": word_info["word"].strip(),
                    "start": word_info["start"],
                    "end": word_info["end"]
                })

    with open(output_path, "w") as f:
        json.dump(words, f, indent=2)

    return words

Detecting Specific Words

def find_words(transcription, target_words):
    """
    Find specific words in transcription with their timestamps.

    Args:
        transcription: List of word dicts with 'word', 'start', 'end'
        target_words: Set of words to find (lowercase)

    Returns:
        List of matches with word and timestamp
    """
    matches = []
    target_lower = {w.lower() for w in target_words}

    for item in transcription:
        word = item["word"].lower().strip()
        # Remove punctuation for matching
        clean_word = ''.join(c for c in word if c.isalnum())

        if clean_word in target_lower:
            matches.append({
                "word": clean_word,
                "timestamp": item["start"]
            })

    return matches

Complete Example: Find Filler Words

import whisper
import json

# Filler words to detect
FILLER_WORDS = {
    "um", "uh", "hum", "hmm", "mhm",
    "like", "so", "well", "yeah", "okay",
    "basically", "actually", "literally"
}

def detect_fillers(audio_path, output_path):
    # Load tiny model (fast!)
    model = whisper.load_model("tiny")

    # Transcribe
    result = model.transcribe(audio_path, word_timestamps=True, language="en")

    # Find fillers
    fillers = []
    for segment in result["segments"]:
        for word_info in segment.get("words", []):
            word = word_info["word"].lower().strip()
            clean = ''.join(c for c in word if c.isalnum())

            if clean in FILLER_WORDS:
                fillers.append({
                    "word": clean,
                    "timestamp": round(word_info["start"], 2)
                })

    with open(output_path, "w") as f:
        json.dump(fillers, f, indent=2)

    return fillers

# Usage
detect_fillers("/root/input.mp4", "/root/annotations.json")

Audio Extraction (if needed)

Whisper can process video files directly, but for cleaner results:

# Extract audio as 16kHz mono WAV
ffmpeg -i input.mp4 -vn -acodec pcm_s16le -ar 16000 -ac 1 audio.wav

Multi-Word Phrases

For detecting phrases like "you know" or "I mean":

def find_phrases(transcription, phrases):
    """Find multi-word phrases in transcription."""
    matches = []
    words = [w["word"].lower().strip() for w in transcription]

    for phrase in phrases:
        phrase_words = phrase.lower().split()
        phrase_len = len(phrase_words)

        for i in range(len(words) - phrase_len + 1):
            if words[i:i+phrase_len] == phrase_words:
                matches.append({
                    "word": phrase,
                    "timestamp": transcription[i]["start"]
                })

    return matches

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: benchflow-ai
• Source: benchflow-ai/skillsbench
• License: Apache-2.0
• Homepage: https://www.skillsbench.ai

Install and usage instructions live in the source repository linked above.