A Deepgram alternative for Hindi call transcription

How do Deepgram and Voice Infra differ?

Deepgram is a speech-to-text company first. Deepgram says its Nova models support 45+ languages, Hindi among them, with a multilingual mode for audio that switches between languages, and Flux adds turn detection for voice agents. Prices are published per minute for every model and add-on, the free credit is generous, and the streaming API handles high concurrency on the self-serve plan. It also sells text to speech and a voice agent API, though Aura’s language list today has no Hindi.

We started from the other end: Indian phone calls. Mira STT is the recognition model inside our voice agents, a fine-tune of the open SraVaani-1.0 model from ARTPARK at IISc, served on our own engine. For recordings, Mira Transcribe runs several recognition passes and a language model that writes one transcript, using any context and vocabulary you send to settle names. A draft is readable within seconds, and the finished transcript labels who spoke and when.

What does an hour of audio cost?

Deepgram’s pre-recorded Nova-3 Monolingual rate is $0.0043 a minute and Multilingual is $0.0052, with speaker diarization charged as an add-on. Its streaming rates are on a limited-time promotion: $0.0048 a minute for Nova-3 Monolingual against a regular $0.0077.

Our recorded-call transcription is ₹18 per hour of audio, which is ₹0.30 a minute, and speaker labels are included. A five-minute call costs ₹1.50. The charge is taken once, when the job completes, and a job that fails or comes back degraded is free.

How does Hindi accuracy compare?

In our own report on RinggAI’s public Hindi benchmark, 10,000 utterances across six corpora, Mira STT scored 9.12% normalised word error rate, the lowest of the five systems. The Deepgram output bundled with that dataset scored 15.79% under the same references and normaliser. Read that with its limits. The vendor outputs are the ones Ringg bundled, so their model versions are Ringg’s choice and may be older than what Deepgram serves today. The open model we fine-tuned was trained on the training splits of those corpora, which helps us on their test splits. Mira STT is not the best on every subset, and the benchmark is read and prompted speech, so it says nothing about phone-call accuracy. Mira Transcribe also adds passes and a language model on top of the recogniser, so the score describes the model, not the product.

The fair test is your own audio. Send the same twenty recorded calls to both and score them against transcripts a Hindi speaker has corrected.

What does switching take?

For recorded calls, upload the file to POST /v2/transcriptions with a line of context and the names you want spelled right, or send an https link for files over 45 MB, then read the job until it completes. Recordings can run to 2 hours. The transcription guide has Python and Node.js examples, and the Hindi STT comparison lists what the other Hindi recognisers cost.

If you were going to pair Deepgram with a separate voice and model to build an agent, our agents bundle all three: ₹1 a minute on T1 with a ₹1,00,000 monthly commitment, ₹3 on T3 with ₹50,000 a month, or ₹5 pay as you go.