Best way seriesNovember 16, 2026
The best way to transcribe audio with AI
Use deepgram/transcribe at 1¢ per declared minute for calls, meetings and video. For Indian-language clips under 30 seconds, use sarvam/transcribe at 2¢.
The best way to transcribe a recording with an AI agent is one call to deepgram/transcribe, which takes a link to an audio or video file and costs 1¢ per minute of audio you declare. For short clips in Hindi, Tamil and other Indian languages, use sarvam/transcribe at a flat 2¢.
What "best" means for transcription
- It accepts the recording you have. A call, a meeting, a voice note or the audio track of a video, a few seconds or a few hours long.
- The cost is known first. A price per minute lets the agent work out the bill from the length before it spends anything.
- The output fits the next step. Plain text is enough for a summary. A meeting needs speaker labels, and captions need timestamps.
- It handles the language spoken. Short clips in Indian languages have their own action.
The options in the Vaaya catalog
Prices are as listed on 16 November 2026. The catalog has current prices.
| Tool | What it returns | Price per call | When to pick it |
|---|---|---|---|
deepgram/transcribe |
A transcript of an audio or video file (Deepgram Nova-3), with optional speaker labels and timestamped segments | 1¢ per declared minute, rounded up, 2¢ minimum | The default: calls, podcasts, voicemails, interviews, screen recordings |
sarvam/transcribe |
{ transcript, language_code, language_probability } (Sarvam Saarika), with optional word timing |
2¢ flat | Clips under about 30 seconds in Indian languages or Indian English |
sarvam/translate |
{ translated_text }, between English and 10 Indian languages |
2¢ for up to 1,000 characters | An English copy of an Indian-language transcript |
files/upload |
A put_url for the file, a file_id and a get_url |
1¢ | Staging a local recording so a transcription call can fetch it |
The default path
Send the recording to deepgram/transcribe unless it is a short clip in an Indian language.
How it is priced. You pass duration_minutes, the approximate length of the recording, and the call is priced on that number: 1¢ per minute, rounded up to a whole minute, with a 2¢ minimum. The parameter accepts 1 to 180, so one call costs at most $1.80. Leave it out and the call is priced as 10 minutes. Declare the real length. Options such as speaker labels do not change the price.
How the audio is supplied. Both transcription actions take a url to a publicly reachable file. For a file on your machine, call files/upload (1¢, up to 100MB per file), send the bytes to the put_url it returns, and pass the get_url as url. The Deepgram action also accepts a file staged with fal/upload (1¢).
When to step up. Add diarize: true for speaker labels and utterances: true for timestamped segments. Set language to an ISO code such as en or hi, or pass detect_language: true.
Indian languages. For a voice note, an IVR reply or a call snippet under about 30 seconds, use sarvam/transcribe. It detects the language, or you can pin it with language_code, for example hi-IN. For longer recordings its catalog description points back to deepgram/transcribe. If your agent works in English, sarvam/translate turns the transcript into English at 2¢ per call, up to 1,000 characters a call.
A worked example: a 47-minute sales call
The numbers here are arithmetic at listed prices on 16 November 2026.
The recording is call.mp3 on your laptop: 46 minutes 20 seconds, two speakers. The agent stages it first:
{
"service": "files",
"action": "upload",
"params": { "filename": "call.mp3", "size_bytes": 44480000, "mime": "audio/mpeg" },
"max_cost_cents": 1
}
It sends the file to the returned put_url, then transcribes from the get_url:
{
"service": "deepgram",
"action": "transcribe",
"params": {
"url": "<get_url from files/upload>",
"duration_minutes": 47,
"diarize": true
},
"max_cost_cents": 47
}
- Staging: 1¢.
- Transcription: 46 minutes 20 seconds rounds up to 47 minutes, and 47 × 1¢ = 47¢.
- Total: 48¢.
Ten calls like this in a week come to 10 × 48¢ = $4.80. A 20-second Hindi voice note is 1¢ to stage and 2¢ on sarvam/transcribe, so 3¢, or 5¢ with one sarvam/translate call for an English copy.
With max_cost_cents set to 20 on the transcription, the call is refused. The response carries the 47¢ price and nothing is charged.
How to run it
Connect your agent to Vaaya with the install guide. Then ask in plain English:
Transcribe
call.mp3with Vaaya. It is 46 minutes 20 seconds long with two speakers, so declare 47 minutes and turn on speaker labels. Spend at most 50 cents in total. Give me the transcript with speaker labels, then the decisions we made.
The agent can ask Vaaya's consult tool to plan the calls. It returns each call with its params and price. Each call carries max_cost_cents, so a price above the number you allowed is refused before anything runs. If you restrict the agent's key by category, transcription sits under Media.
What to check, and where this falls short
- Find the text. The transcript is at
results.channels[0].alternatives[0].transcript. - Compare lengths.
metadata.durationis the real duration in seconds. If it is far from what you declared, the agent fetched the wrong file or misjudged the length. - Check what matters. Punctuation, numbers and dates are formatted automatically. Read names, figures and dates against the recording before an agent acts on them.
- Name the speakers yourself.
diarize: trueadds speaker labels. Matching a label to a person is your job. - Read the language fields.
sarvam/transcribereturnslanguage_codeandlanguage_probability. Check both before you trust a detected language, and pinlanguage_codewhen you know it.
The limits:
- Both actions work on recorded files. The catalog has no live or streaming transcription action.
- One
deepgram/transcribecall covers at most 180 declared minutes. Split longer recordings. sarvam/transcribeis for clips under about 30 seconds and files up to 25MB.sarvam/translatetakes 1,000 characters per call, so chunk long transcripts by paragraph.- A staged
get_urllasts about an hour.files/getreturns a fresh one at no charge. - For Deepgram, the catalog gives
enandhias example language codes and no full list. For a long recording in another language, test a two-minute section first: 2¢ plus staging.
A failed call is not charged. A call that completes with a poor transcript is charged, and you can raise a dispute on any charge from Transactions in your account. A person at Vaaya reviews it.
Going the other way, from a script to audio, is covered in how to create a voiceover for a script.
Why transcription suits a pay-per-call wallet
Recordings arrive unevenly: nothing for a week, then a day of back-to-back calls. A price per minute with no subscription follows that, and Deepgram and Sarvam both draw on one balance with no separate account or API key for each. A new personal account starts with $2 of welcome credit, which is 200 declared minutes at 1¢ a minute (pricing). For a batch job, a daily or weekly limit on the agent's key (spending controls) stops a folder of recordings from running past the budget.
Questions
How much does it cost to transcribe audio with AI?
Through Vaaya, deepgram/transcribe costs 1¢ per declared minute of audio, rounded up, with a 2¢ minimum, so a 47-minute call is 47¢. sarvam/transcribe is a flat 2¢ per short clip. Prices are as listed on 16 November 2026.
Can an AI agent transcribe a video file?
Yes. deepgram/transcribe takes a URL to an audio or video file and transcribes the audio track. A file on your own machine has to be staged first, for example with files/upload at 1¢.
How do I transcribe Hindi, Tamil or other Indian-language audio?
For clips under about 30 seconds, use sarvam/transcribe at a flat 2¢. It detects the language, or you can pin a code such as hi-IN. For longer recordings the catalog points to deepgram/transcribe, which takes a language code such as hi.
Can I get speaker labels and timestamps in the transcript?
Yes. On deepgram/transcribe, diarize: true adds speaker labels and utterances: true returns timestamped segments, at the same price. On sarvam/transcribe, with_timestamps: true adds word timing.