whisper

Whisper

Whisper (OpenAI) is the industry standard for Speech-to-Text. It supports 99 languages and translation. V3 (large-v3) is the current state of the art.

When to Use

• Speech-to-Text Audio Transcription: Transcribing multi-lingual audio recordings into punctuated, accurate text.
• Audio Translation to English: Automatically translating non-English speech directly into English transcripts.
• Word-Level Subtitles & Timestamps: Generating SRT and VTT subtitles with word-level alignment for video editors.
• High-Throughput Inference with Faster-Whisper: Running CTranslate2-accelerated inference with 4x speedup and lower VRAM.

Quick Start

import whisper

# Load pre-trained model and transcribe audio file
model = whisper.load_model("base")
result = model.transcribe("interview.mp3")

print("Detected language:", result["language"]) print("Full Transcription:\n", result["text"])

## Core Concepts

### High-Speed Audio Transcription with Faster-Whisper

Using CTranslate2 acceleration for production-grade transcription:

```python
from faster_whisper import WhisperModel
import time

# Load model in INT8/FP16 precision on GPU
model = WhisperModel(
    model_size_or_path="large-v3",
    device="cuda",
    compute_type="float16"
)

# Transcribe audio file with voice activity detection (VAD)
start_time = time.time()
segments, info = model.transcribe(
    "interview_audio.mp3",
    beam_size=5,
    language="en",
    vad_filter=True, # Skip silent audio gaps
    vad_parameters=dict(min_silence_duration_ms=500)
)

print(f"Detected language: {info.language} with probability {info.language_probability:.2f}")

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

print(f"Transcription completed in {time.time() - start_time:.2f} seconds.")

Generating Word-Level Timestamps for Subtitles

Extracting exact timestamps for every individual spoken word:

segments, info = model.transcribe(
    "keynote_speech.mp4",
    word_timestamps=True
)

for segment in segments:
    for word in segment.words:
        print(f"Word: {word.word:12s} ({word.start:.2f}s - {word.end:.2f}s, prob: {word.probability:.2f})")

Direct Audio Translation to English

Transcribing foreign audio directly into English text:

# Task = 'translate' translates non-English speech to English
segments, info = model.transcribe(
    "spanish_lecture.mp3",
    task="translate",
    beam_size=5
)

full_english_transcript = " ".join([segment.text for segment in segments])
print("Translated Transcript:\n", full_english_transcript)

Common Patterns

Word-Level Timestamps with faster-whisper

Problem: Standard Whisper model runs slowly on CPU and only returns coarse segment timestamps.

Solution: Use CTranslate2-accelerated faster-whisper with word-level timestamps:

from faster_whisper import WhisperModel

# 4x faster execution with 8-bit quantization on GPU or CPU
model = WhisperModel("small", device="auto", compute_type="int8")

segments, info = model.transcribe("speech.wav", word_timestamps=True)

for segment in segments:
    for word in segment.words:
        print(f"[{word.start:.2f}s -> {word.end:.2f}s] {word.word}")

Best Practices

Do:

• Target faster-whisper with compute_type="float16" or "int8_float16" for up to 4x throughput and lower VRAM.
• Enable vad_filter=True to filter out background silence and prevent hallucinated repetitive loops.
• Specify language explicitly when known in advance to bypass language identification overhead.
• Use whisper-large-v3 or whisper-large-v3-turbo for optimal transcription accuracy.

Don't:

• Process long multi-hour audio files as a single unbuffered stream without VAD chunking.
• Use standard OpenAI whisper Python package in production without checking if faster-whisper provides better throughput.
• Feed extremely noisy audio without applying pre-processing bandpass filters or noise suppression.

Troubleshooting

Error	Cause	Solution
FileNotFoundError: [Errno 2] No such file or directory: 'ffmpeg'	System dependency ffmpeg not installed in PATH.	Install ffmpeg: brew install ffmpeg or apt-get install ffmpeg.
CUDA Out of Memory loading 'large-v3'	Large model weights exceed available GPU VRAM.	Use compute_type="float16" or switch to medium / small model size.
Transcription hallucinates repetition loops	Audio contains prolonged silence or background music.	Set no_speech_threshold=0.6 and condition_on_previous_text=False.

References

• Whisper GitHub