An Hour of Audio Now Costs a Dime. Here's What to Build on That.

Microsoft is currently charging ten cents to transcribe an hour of audio. Meta wants eighteen cents. Google’s per-token billing works out to roughly thirty cents for pre-recorded files. A month ago, those numbers would have looked like a pricing error. Two years ago, anything under a dollar per audio hour with professional-grade accuracy was considered a procurement win.

Three speech-to-text models shipped within nine days of each other: Google’s Gemini 3.5 Transcribe on August 26, Meta’s Muse Voice Transcribe on September 1, and Microsoft’s MAI-Transcribe-2 on September 3. DeepLearning.AI’s newsletter The Batch has the full launch coverage, and it’s worth reading. But the part that matters for content and marketing teams is shorter. The transcript is close to free now. The value moved to whatever you build on top of it.

What shipped, in numbers

All three land under 4 percent word error rate on Artificial Analysis’ speech benchmark, which is the zone where accuracy stops being a differentiator and everything else starts mattering more.

One architecture detail from Meta stuck with me. Muse Voice Transcribe processes audio in 80-millisecond chunks, and at each chunk the model either emits text or predicts a “keep listening” token and gathers more context. Meta calls this adaptive delay, and the speed-accuracy tradeoff is learned during training through an accuracy reward and a latency penalty rather than configured by an engineer. Easy speech transcribes fast; ambiguous words get extra context. It’s a small thing, but it tells you where this category is heading: the models are making judgment calls that used to be configuration options.

Why the price collapsed

The Batch’s framing is that transcription is “ruthless” on cost because the real prize is the voice-agent market, where AI systems still compete against human agents staffing phones. Whoever owns cheap, accurate speech-to-text owns the substrate under every voice agent. Content producers are not the target of this war. We just get to stand downstream of it.

That’s a nice position to be in. The recordings you already have, the webinars and podcasts and interview files sitting unreviewed in a drive folder somewhere, just became the cheapest raw material you own.

The pipeline: one recording, five assets

Here’s the workflow I’d set up this week. It assumes one hour of source audio (a webinar, a podcast episode, a customer interview) and turns it into a week of content.

Step 1: Transcribe for pennies. Pick a model based on the guide in the next section. You’ll spend $0.10 to $0.30 for the hour. Ask for diarization if your recording has multiple speakers.

Step 2: Show notes and a summary post. Feed the transcript to your writing model with a prompt that asks for a structured summary, the five strongest moments, and their approximate timestamps. Edit for 15 minutes. Publish.

Step 3: Social quotes. Same transcript, different prompt: pull the five best one-to-two-sentence claims with speaker attribution. These are your posts for the week, sourced from something a real person actually said, which beats invented LinkedIn wisdom on both ethics and performance.

Step 4: A newsletter section. Condense the takeaways into 150 words. If you record regularly, this becomes a recurring column that writes itself from material you produced anyway.

Step 5: An FAQ page. Audience questions tend to repeat across interviews and Q&A portions. Extract every question asked, group the duplicates, and answer the top ones on a single page. This is the least glamorous asset and possibly the most durable one for search.

If you want a sixth asset, Google’s word-level timestamps make caption-aligned clip extraction scriptable, so short video cuts stop requiring manual scrubbing.

Run the arithmetic on a realistic month: 10 hours of audio at $0.30 is $3.00. At Microsoft’s promo price it’s a dollar. Add the LLM passes over the transcripts and you’re still under the cost of a single freelance blog post, producing five asset types from material you already recorded.

Which model to pick

• Bulk pre-recorded audio, cost above all: MAI-Transcribe-2. Best accuracy of the three, absurd throughput (an hour of audio in 10 seconds), and the $0.10 promo price through the end of 2026.
• Live events, panels, anything streaming: Muse Voice Transcribe. Best streaming accuracy at 3.1 percent, and 20+ speaker diarization covers conference scenarios the other two can’t.
• Multilingual content or clip production: Gemini 3.5 Transcribe. The 85+ language count and word-level timestamps are doing work the others don’t.

One caution: prices in this category are moving weekly. Microsoft’s rate is explicitly promotional. Whichever you choose, re-check pricing before you commit a year of archive to it, and keep the other two on file as leverage when you renegotiate with whatever vendor you’re currently paying ten times more.

What to do now

1. List the recordings your team made last quarter that nobody repurposed.
2. Run one hour through two of these models and compare the output against each other, not against perfection.
3. Write the single prompt that converts your transcript into show notes, and save it as a template. That prompt is now an asset.
4. Track hours saved per recording for a month. That number is your business case.

The uncomfortable version of this advice: if transcription cost was ever your reason for not repurposing recordings, that reason expired this month. The bottleneck is no longer the transcript. It’s the editorial hour you spend turning it into something worth reading, and no price war is coming for that.