Streaming Transcription Basics

Understanding Streaming Transcription

Welcome back! In the last lesson, you learned how to use system prompts to control the format of your audio transcriptions, making it easy to generate meeting notes or action item lists from the same recording. Now, we'll build on that foundation by exploring streaming transcription . This lesson will help you understand how streaming works and when you might want to use it in real-world applications.

What You'll Learn

In this lesson, you will: Understand what streaming transcription is and how it works. Learn how to simulate streaming transcription with the Whisper API . See how to implement streaming transcription in Python . Know when to choose streaming transcription for your applications. By the end, you'll know how to implement streaming transcription in your own projects and when it's the right choice.

What is Streaming Transcription?

Streaming transcription breaks the audio into smaller pieces ("chunks") and sends them to the API as they become available. The API returns partial results as soon as they're ready, so you can display the transcript in real time, even as the audio is still being processed. Think of it like watching a live sports game versus watching highlights later. With streaming, you get the action as it happens, chunk by chunk. This creates a much more interactive and responsive user experience.

When to Use Streaming Transcription

Streaming transcription is ideal when: You have long audio files - Instead of waiting 30 seconds for a 10-minute recording, users see results immediately You're building interactive applications - Live captioning, voice assistants, or real-time meeting transcription User experience matters - When you want your app to feel fast and responsive You're processing live audio - Webinars, phone calls, or live events where audio is still being generated You want to show progress - Users can see that something is happening rather than staring at a loading screen You might stick with regular batch processing when you're processing short files in the background, doing bulk processing, or when the user doesn't need to see results immediately.

Join the 1M+ learners on CodeSignal

Be a part of our community of 1M+ users who develop and demonstrate their skills on CodeSignal

Implementing Streaming Transcription

Now, let's see how the streaming transcription function works. Note: OpenAI's Whisper API doesn't currently support real streaming, so we'll simulate the streaming behavior to demonstrate the concept:

Python

@retry_on_error(max_retries=3, delay=5)
def transcribe_audio(audio_file_path: str, prompt: Optional[str] = None, stream: bool = False):
    try:
        with open(audio_file_path, "rb") as audio:
            if stream:
                # Since OpenAI Whisper API doesn't support streaming, we'll simulate it
                response = client.audio.transcriptions.create(
                    model="gpt-4o-transcribe",
                    file=audio,
                    prompt=prompt or "",
                    response_format="text",
                )
                
                # Simulate streaming by yielding text in chunks
                def stream_generator():
                    text = response if isinstance(response, str) else response.text
                    # Split text into words for streaming simulation
                    words = text.split()
                    chunk_size = max(1, len(words) // 10)  # Aim for ~10 chunks
                    
                    for i in range(0, len(words), chunk_size):
                        chunk = " ".join(words[i:i + chunk_size])
                        if i > 0:
                            chunk = " " + chunk
                        yield chunk
                        time.sleep(0.3)  # Simulate processing delay
                
                return stream_generator()
            else:
                response = client.audio.transcriptions.create(
                    model="gpt-4o-transcribe",
                    file=audio,
                    prompt=prompt or "",
                    response_format="text",
                )
                return response if isinstance(response, str) else response.text
    except Exception as e:
        print(f"Error during transcription: {e}")
        raise

@retry_on_error(max_retries=3, delay=5)
def transcribe_audio(audio_file_path: str, prompt: Optional[str] = None, stream: bool = False):
    try:
        with open(audio_file_path, "rb") as audio:
            if stream:
                # Since OpenAI Whisper API doesn't support streaming, we'll simulate it
                response = client.audio.transcriptions.create(
                    model="gpt-4o-transcribe",
                    file=audio,
                    prompt=prompt or "",
                    response_format="text",
                )
                
                # Simulate streaming by yielding text in chunks
                def stream_generator():
                    text = response if isinstance(response, str) else response.text
                    # Split text into words for streaming simulation
                    words = text.split()
                    chunk_size = max(1, len(words) // 10)  # Aim for ~10 chunks
                    
                    for i in range(0, len(words), chunk_size):
                        chunk = " ".join(words[i:i + chunk_size])
                        if i > 0:
                            chunk = " " + chunk
                        yield chunk
                        time.sleep(0.3)  # Simulate processing delay
                
                return stream_generator()
            else:
                response = client.audio.transcriptions.create(
                    model="gpt-4o-transcribe",
                    file=audio,
                    prompt=prompt or "",
                    response_format="text",
                )
                return response if isinstance(response, str) else response.text
    except Exception as e:
        print(f"Error during transcription: {e}")
        raise

The key difference here is the stream parameter. When set to True, the function simulates streaming by:

Getting the complete transcription from the Whisper API
Breaking it into chunks by splitting the text into words and grouping them
Yielding chunks progressively with artificial delays to simulate real-time processing
Returning a generator that yields each chunk as it becomes "available"

We also have a convenience function that makes it easy to choose between batch and streaming modes: