Transcription and speech synthesis are priced differently from chat: per minute of audio or per thousand characters, not per token. That makes them easy to under-budget. This post turns one vendor catalog into plain rupee figures for an hour of audio and a hundred thousand characters, and flags where the data is odd.
Disclosure: AICredits links on this site are affiliate links. I read the public catalog on 2026-10-16 and made no authenticated calls. I did not test accuracy or latency of any model, so this is a price post only.
Where is speech pricing in the catalog?
Not in the per-token fields. Audio and text-to-speech entries have zero token prices and carry their pricing in a JSON string in the field named image_cost_matrix (the name is a leftover from image models). Formulas I saw: per_minute for transcription, per_character for TTS, and flat for music generation. Some entries include a source URL and a note.
What does an hour of transcription cost?
I parsed those fields and converted USD to INR at the catalog's own exchange rate (95.10014), showing a second figure with the 1.05 factor that token-priced entries carry, which is an assumption for these entries:
Speech-to-text (cost for 1 hour of audio):
mistralai/voxtral-small-24b-2507 Rs 17.12/h at rate 95.10014 | with x1.05: Rs 17.97/h (USD 0.003/min)
openai/whisper-1 Rs 34.24/h at rate 95.10014 | with x1.05: Rs 35.95/h (USD 0.006/min)
sarvam/saarika-v2 Rs 37.50/h as listed in INR (0.625/min)
sarvam/saarika-v2.5 Rs 37.50/h as listed in INR (0.625/min)
The Whisper entry cites OpenAI's pricing page as its source; I did not re-verify that. The Sarvam entries list a note saying the price is the midpoint of a Rs 30 to 45 per hour range, billed per second, citing Sarvam's pricing documentation. So the Rs 37.50 is the catalog maintainer's midpoint, not a flat vendor rate.
Quick scale check: transcribing 100 hours of call recordings would be about Rs 1,712 on Voxtral, Rs 3,424 on Whisper, or Rs 3,750 on Saarika, before any markup. Compared with the cost of reviewing calls by hand, the audio-to-text step is rarely the expensive part. The language model that reads the transcript usually is.
What does text-to-speech cost?
Text-to-speech (cost for 100,000 characters):
sarvam/bulbul-v2 Rs 150.00 (1.5/1k chars; INR 15 per 10K chars = 1.5 per 1K)
sarvam/bulbul-v3 Rs 300.00 (3.0/1k chars; INR 30 per 10K chars = 3.0 per 1K)
A rough conversion for planning: a typical English page is around 3,000 characters, so 100,000 characters is about 33 pages of narration. At Rs 3.00 per thousand characters, one page costs about Rs 9. That figure is the only one I trust here, because the Bulbul v3 rate matches what Sarvam's own pricing page listed for pay-as-you-go on the same day (Rs 3.00 per 1,000 characters).
Where does the catalog disagree with the vendor?
Three things I noticed:
- The gateway lags the vendor. Sarvam's own pricing page that day listed Saaras v3 speech to text at Rs 0.49 per minute and Bulbul V4 Flash at Rs 2.00 per 1,000 characters. The catalog listed only Saarika v2 and v2.5 and Bulbul v2 and v3. If you want the newest Sarvam models, check whether the gateway has them before you plan around it. My Sarvam API guide covers the vendor's own lineup and prices.
- Music is labelled as transcription. The two Google Lyria music-generation entries (flat USD 0.04 per clip and USD 0.08 per song) appear in
audiomode withsupported_apisofaudio_transcriptions. Music generation is not transcription, so filter by model name, not only by mode, if you build a picker on this data. - Price units differ. Per minute, per hour, per character, per clip. Convert to one unit before you compare, as the script above does for transcription (per hour) and TTS (per 100,000 characters).
How do you decide between them?
Price is the last filter, not the first.
- Accuracy on your audio. Indian-accented English, code-mixed speech and noisy call audio behave differently on different models. Use the speech-to-text test protocol for your own recordings before choosing.
- Language coverage. Check the model's supported languages on the vendor's page. The catalog does not list them.
- Latency and streaming. Voice agents need streaming transcription and quick speech output. The catalog's
audio_transcriptionsandaudio_speechAPIs say nothing about real-time support. See the voice agent latency budget. - Total pipeline cost. For a phone agent, transcription and speech are two of three costs; the model is the third. The voice agent cost per minute calculator puts them together.
The AICredits signup is an affiliate link. For the data behind the rest of the catalog, see the catalog analysis.
What I could not verify
- Transcription accuracy, latency or reliability of any model listed.
- That OpenAI's and Mistral's USD prices in the catalog match their own pages today (the Whisper entry cites OpenAI's page; I did not re-check it).
- That AICredits charges the converted amounts shown here.
- Whether the gateway supports real-time speech through any of these entries.
