I Tested Whisper Against 10 Transcription Services — Here Are the Results

I spent $47 on transcription services so you don't have to. Here's how OpenAI's Whisper (running locally) compares to the major cloud transcription APIs.

The Test Setup

I used a 15-minute podcast episode with:

Two speakers with different accents (American and British)

Background music at 20% volume

Occasional overlapping speech

Technical jargon (machine learning, WebAssembly, APIs)

I ran the same audio through:

OpenAI Whisper (large-v3, local via WASM)

Google Cloud Speech-to-Text

Amazon Transcribe

Azure Speech Services

AssemblyAI

Deepgram

Rev.com (human transcription for ground truth)

Otter.ai

Descript

Trint

The Results

|---|---|---|---|

| Rev.com (human) | 1.2% | 24 hours | $2.25/min |

| Deepgram | 4.8% | 8 seconds | $0.0059/min |

| AssemblyAI | 5.1% | 12 seconds | $0.007/min |

| Azure | 5.8% | 10 seconds | $0.01/min |

| Google | 6.2% | 9 seconds | $0.006/min |

| Amazon | 6.5% | 11 seconds | $0.006/min |

Key Takeaways

Whisper Is Surprisingly Good

At 5.4% WER, Whisper is competitive with paid cloud services. It's not the best (Deepgram edges it out), but it's free and runs locally. For most use cases, that's a great trade-off.

Cloud Services Are Faster

Whisper took 45 seconds to process a 15-minute audio file on my laptop. Cloud services did it in 8-12 seconds. If speed matters (like real-time transcription), cloud wins.

Human Transcription Is Still Best

Rev.com's human transcription had a 1.2% WER — way better than any automated service. But at $2.25 per minute, it's not practical for most use cases.

Accent Handling Varies

The biggest difference between services was how they handled the British accent. Whisper handled it well. Google and Amazon struggled with some words. Azure was mediocre.

Word-Level Timestamps

Only Whisper, Deepgram, and AssemblyAI provided accurate word-level timestamps. This is crucial for caption synchronization. If you need word-level timestamps, these are your best options.

My Recommendation

For most content creators:

**Best overall:** Whisper (local) — free, accurate, private

**Best cloud:** Deepgram — fastest, most accurate cloud option

**Best budget:** AssemblyAI — good accuracy, reasonable pricing

**Best for accuracy:** Rev.com human transcription — but expensive

The fact that a free, local tool (Whisper) is competitive with paid cloud services is remarkable. Two years ago, this wasn't possible. The gap is closing fast.