Transcribing 400 hours of audio costs $60 a month on AssemblyAI Universal-2 and $1,308 on ElevenLabs Scribe v2. Same audio, same month, twenty-two times the invoice. Neither vendor is being dishonest — $0.0025 and $0.0545 per minute are both published, both correct — but almost nobody notices the gap, because per-minute prices with four decimal places do not register as different in the way $60 and $1,308 do. That translation is what this calculator does.
Two units that cannot be compared by eye
Speech AI bills on two incompatible bases, and the calculator handles each separately before adding them together.
Speech-to-text bills per minute of audio. The unit is wall-clock duration of the file you submit. Not words recognised, not speakers detected, not characters returned. A 60-minute recording is 60 billable minutes whether it contains a dense panel discussion or one person saying "yes" and then nothing.
Text-to-speech bills per character of input. Characters, not words — and characters include spaces, punctuation, and in most implementations any SSML markup you send. ElevenLabs is the exception in this index, billing $0.17 per minute of generated audio on its Pro tier, which is why the calculator converts everything to a per-minute figure before ranking.
Two conversions make the units meet. Synthesis input is estimated at 5.7 characters per word, which is the standard English average once spaces and punctuation are counted. Delivery is estimated at 150 spoken words per minute, the long-standing broadcast pace — audiobook narration runs slower at 140 to 155, conversational agents faster at 160 to 180. Both are averages you can override by reading your own logs, and both are close enough that they will not change which vendor wins.
Working one month through
Take the defaults: 400 hours of audio to transcribe, 250,000 words to synthesise.
Transcription. 400 hours × 60 = 24,000 minutes. On Deepgram Nova-3 at $0.0048 a minute: 24,000 × $0.0048 = $115.20. Divide the per-minute rate by nothing and multiply by 60 and you get the number that makes vendors comparable — $0.288 per hour of audio.
Synthesis. 250,000 words × 5.7 = 1,425,000 characters. On Google TTS Neural2 at $16.00 per million characters: 1.425 × $16.00 = $22.80. Those words play for 250,000 ÷ 150 = 1,667 minutes, so the effective rate is $0.0137 per minute of finished speech.
The month is $138.00, and the striking part is the split: transcription is five times the synthesis bill even though the synthesis output runs for 28 hours against 400 hours transcribed. Transcription is the expensive half of most speech pipelines and the half teams pay least attention to.
Cost per audio hour is the number to quote
Per-minute pricing hides magnitude. Per-hour pricing does not, and audio arrives in hours — meetings, calls, interviews, episodes. Here is the whole speech-to-text index at that framing, with the monthly bill for 400 hours:
| Model | $/min | $/audio hour | 400 hours |
|---|---|---|---|
| AssemblyAI Universal-2 | 0.0025 | $0.15 | $60.00 |
| OpenAI gpt-4o-mini-transcribe | 0.0030 | $0.18 | $72.00 |
| Google STT V2 dynamic batch | 0.0030 | $0.18 | $72.00 |
| AssemblyAI Universal-3.5 Pro | 0.0035 | $0.21 | $84.00 |
| OpenAI gpt-transcribe | 0.0045 | $0.27 | $108.00 |
| Deepgram Nova-3 | 0.0048 | $0.288 | $115.20 |
| Deepgram Nova-3 multilingual | 0.0058 | $0.348 | $139.20 |
| OpenAI gpt-4o-transcribe | 0.0060 | $0.36 | $144.00 |
| Google STT V2 standard | 0.0160 | $0.96 | $384.00 |
| ElevenLabs Scribe v2 (Pro) | 0.0545 | $3.27 | $1,308.00 |
Note Google's own two rows: $0.016 a minute for standard synchronous recognition against $0.003 for dynamic batch. That is a 5.3-fold discount for accepting a queue instead of an immediate answer. If your workload is recorded rather than live — and most transcription is — the batch tier is the same model at $72 a month instead of $384. Asynchronous processing is the single largest saving available on this page, and it costs you nothing but latency.
On the synthesis side the same spread appears: Google Standard/WaveNet at $4 per million characters is $5.70 for our month, Neural2 at $16 is $22.80, Chirp 3 HD and Deepgram Aura-2 at $30 are $42.75 each, Studio at $160 is $228, and ElevenLabs at $0.17 a minute is $283.33. Fifty times, top to bottom, for the same script.
You are paying for silence
Because transcription bills wall-clock duration, every pause, every hold, every stretch of an empty conference bridge is billed at the same rate as speech. A recorded meeting that runs 60 minutes with 15 minutes of dead air, cross-talk before the start and someone forgetting to stop the recorder is 25% waste.
Apply that to our example: strip 400 hours to 300 with voice activity detection at ingest and the Nova-3 bill drops from $115.20 to $86.40. Small in isolation, 25% forever, and it improves output quality too — long silences are where spurious hallucinated text tends to appear in transcripts. Trimming at ingest is a couple of hours of work with an off-the-shelf VAD library and it is the first thing to do on any archive workload.
Where the price difference stops mattering
The entire spread of transcription pricing is $3.12 per audio hour, from AssemblyAI's $0.15 to ElevenLabs' $3.27. Now price accuracy on the same basis. An hour of audio at 150 words a minute contains about 9,000 words, so one percentage point of word error rate is 90 errors per hour of audio. If a human reviews the transcript and each correction takes 15 seconds to find and fix, that point of WER costs 22.5 minutes of someone's time — about $15 at a modest loaded rate, or roughly five times the entire vendor price spread.
So the ranking above only decides your bill when nothing downstream reads the output carefully. Where accuracy earns its premium is specific and predictable:
Accented and code-switched speech, where models trained mostly on North American English degrade sharply and multilingual tiers such as Deepgram Nova-3 multilingual at $0.0058 exist precisely for this. Overlapping speakers, where diarisation quality decides whether a four-person meeting transcript is usable at all — test this with your worst recording, not your best. Domain vocabulary, where drug names, part numbers and legal terms come back as plausible near-misses; most vendors accept custom vocabulary or keyword boosting, and using it is worth more than switching provider. And anything feeding an LLM afterwards, where transcription errors propagate silently into summaries that read fluently and are wrong.
Run your own worst hour of audio through the top three candidates before you pick on price. Then use this calculator to check that the winner is affordable at volume — and see the hidden costs of AI guide for the review time that follows, the price index for sources on every figure above, and the image and video cost calculator if speech is one part of a larger media pipeline.
Prices used by this calculator were last verified on 15 August 2026 from the vendors' own pricing pages. See the full price index for every figure and its source, the change log for what has moved recently, and the methodology for how the index is maintained and where its limits are.