Whisper doesn’t support streaming — is there a roadmap for it?

[Santoshchodipilli]

I have a challenge, that I want to convert speech-to-text using whisper API.

Note : I am expecting the streaming response only.

[MansoobeZahra]

OpenAI's Whisper API (and the open-source model) does not natively support real-time streaming transcription cuz of offline batch processing, not low-latency streaming and API limitations.
If you need near-real-time transcription, here are some approaches:

1. Chunked Processing (Pseudo-Streaming)
   Split Audio into Short Segments (e.g., 5–10 sec chunks) and send them sequentially to Whisper.
   Pros: Works with existing Whisper API.
   Cons: Higher latency, no cross-chunk context.

2. Use Whisper in Local Mode with Buffering
   Run faster-whisper (optimized implementation) locally and stream mic input in chunks.

3. Alternative Streaming ASR Services
   AssemblyAI / Deepgram / Rev.ai offer real-time streaming APIs.
   Google Speech-to-Text has live streaming support.

   OpenAI has not officially announced streaming support for Whisper.
   You can +1 the issue in the Whisper GitHub repo or request it via OpenAI’s support.

[SameerSenapati17]

If you need speech-to-text with Whisper API in streaming mode, you can’t get true “word-by-word” streaming directly from OpenAI’s current Whisper endpoint — it only supports batch transcription.

To achieve streaming behavior, the common and effective approach is:

1. Stream audio chunks from the client (WebSocket or WebRTC).

2. Send each chunk to a lightweight real-time STT service (e.g., Vosk, Deepgram, AssemblyAI) that supports streaming.

3. Optionally pass partial or full segments to Whisper for higher accuracy after the stream completes.

[Bahtya]

Whisper does not support streaming transcription

The Whisper API (/v1/audio/transcriptions) processes the entire audio file at once and returns the complete transcription. There is no streaming mode.

Alternatives for streaming/real-time STT

1. OpenAI Realtime API (recommended)

The Realtime API supports real-time audio streaming with low-latency transcription:

import asyncio
from openai import AsyncOpenAI

# Use the Realtime API via WebSocket
# See: https://platform.openai.com/docs/api-reference/realtime

The Realtime API is designed for live conversations and streams audio in real-time.

2. Use chunked processing with Whisper

If you must use Whisper, you can simulate streaming by sending audio chunks:

import audioop
import io
from openai import OpenAI

client = OpenAI()

async def transcribe_chunks(audio_stream, chunk_duration_ms=5000):
    buffer = b""
    for chunk in audio_stream:
        buffer += chunk
        if len(buffer) >= chunk_duration_ms * 32:  # rough byte estimate
            transcript = client.audio.transcriptions.create(
                model="whisper-1",
                file=("audio.wav", buffer),
                response_format="text",
            )
            yield transcript
            buffer = b""

Note: This loses context between chunks, so words at boundaries may be cut.

3. Google Speech-to-Text or Deepgram

For production streaming STT, consider services that natively support it:

• Google Cloud Speech-to-Text (Streaming API)
• Deepgram (streaming transcription)
• AssemblyAI (real-time transcription)

Summary

OpenAI has not announced plans to add streaming to the Whisper API. The Realtime API is their solution for real-time audio use cases.