Multimodal: Responding to Speech

AI that listens and responds

Two approaches to speech + AI

Traditional speech pipeline vs multimodal audio

Feature	Traditional Pipeline	Multimodal (GPT-4o)
How it works	Speech → Text → LLM → Text → Speech (3 separate steps)	Speech → GPT-4o → Response (direct audio processing)
Services needed	Azure Speech + Azure OpenAI (separate services)	GPT-4o multimodal (one model)
Latency	Higher (multiple API calls)	Lower (single call)
Nuance	Loses tone, emphasis, emotion in transcription	Preserves audio nuance — can understand tone and intent
Best for	When you need the transcript AND the AI response	When you want natural, conversational AI interaction

Using GPT-4o with audio input

import base64

# Read audio file
with open("question.wav", "rb") as f:
    audio_data = base64.b64encode(f.read()).decode()

response = chat.complete(
    model="gpt4o-deployment",
    messages=[
        {"role": "system", "content": "You are a helpful assistant. Respond naturally to spoken questions."},
        {"role": "user", "content": [
            {"type": "input_audio", "input_audio": {"data": audio_data, "format": "wav"}}
        ]}
    ]
)

print(response.choices[0].message.content)

What’s happening:

• The audio file is encoded as base64 and sent directly to GPT-4o
• The model processes the audio natively — no separate speech-to-text step
• The response is text (or can be audio in supported configurations)

DataFlow Corp scenario: DataFlow builds a voice-enabled analytics dashboard. Managers speak queries like “What were our top-selling products last quarter?” GPT-4o understands the spoken question, queries the data, and responds with the answer.

💡 When to use traditional pipeline vs multimodal

Use the traditional pipeline (Speech + LLM) when:

• You need the transcript for records or compliance
• You need custom speech recognition (industry terms, accents)
• You need speech translation between languages
• Budget is tight (dedicated speech service can be cheaper)

Use multimodal (GPT-4o) when:

• You want the simplest possible architecture
• Tone and emotional context matter for the response
• Low latency is critical
• You’re already using GPT-4o for other modalities (text, images)

🎬 Video walkthrough

Flashcards

Question

How can GPT-4o process spoken questions?

Click or press Enter to reveal answer

Answer

GPT-4o is multimodal — it can accept audio as a direct input modality. The audio is encoded as base64 and sent in the messages array. The model processes the audio waveform directly without needing a separate speech-to-text service.

Click to flip back

Question

What is the advantage of multimodal audio over a traditional speech pipeline?

Click or press Enter to reveal answer

Answer

Lower latency (single API call vs three), preserves audio nuance (tone, emphasis), and simpler architecture. The traditional pipeline requires separate Speech-to-text → LLM → Text-to-speech services.

Click to flip back

Question

When should you use the traditional speech pipeline instead of multimodal?

Click or press Enter to reveal answer

Answer

When you need the transcript for records, need custom speech recognition for industry terms, need speech translation, or when budget is tight (dedicated speech service can be cheaper).

Click to flip back

Knowledge Check

Knowledge Check

MediSpark's doctors want to dictate clinical notes and have AI summarise them. They also need the raw transcript saved to patient records. Which approach should they use?

AGPT-4o multimodal — process audio directly for the summaryBAzure Speech-to-text first (save transcript), then GPT-4o for summarisationCGPT-4o for transcription and summarisation in one stepDAzure Speech synthesis — to read back the notes

Check Answer

Knowledge Check

DataFlow Corp wants a voice-enabled dashboard where managers ask spoken questions and get instant answers. Tone of voice should influence the response style. What's the best approach?

AAzure Speech-to-text + text-based LLM — convert speech firstBGPT-4o multimodal — process audio directly to preserve tone and respond naturallyCAzure Speech translation — to translate the question firstDAzure AI Language — for sentiment analysis of the speech

Check Answer

---

Next up: Azure Speech in Foundry Tools — building apps with dedicated speech services.