AI agents · OpenClaw · self-hosting · automation

Quick Answer

Speech-to-Text API Price per Minute: Streaming & Webhooks

Published:

The short answer

For recorded audio, ElevenLabs Scribe v3 and AssemblyAI-STT-3.7 are the cheapest multilingual APIs at about $0.0033–$0.0035 per minute; for streaming, ElevenLabs Scribe v3 Live and Deepgram Nova-3 lead on price; Gladia is cheapest of all if your audio is in one of Solaria-3’s five languages. Every vendor below supports webhooks (callbacks) for asynchronous jobs. Prices are pay-as-you-go list rates read from each vendor’s pricing page on November 18, 2026, converted to per minute.

API (model)Batch / pre-recordedStreamingWebhook / callbackFree creditWatch out for
ElevenLabs Scribe v3$0.20/hr ≈ $0.0033/minScribe v3 Live $0.22/hr ≈ $0.0037/minwebhook=true + webhook_metadataFree API tiertranscript_edit adds 30%
AssemblyAI-STT-3.7$0.21/hr = $0.0035/minRealtime $0.30/hr ($0.005/min) · Realtime-Agent $0.45/hr ($0.0075/min)webhook_url$50Streaming billed by session time
Speechmatics Melia 1$0.24/hr = $0.004/min$0.40/hr ($0.0067/min) · Linden 1 (voice agents) $0.30/hrBatch notifications$100Allowlist their egress IPs
Deepgram Nova-3$0.0043/min$0.0048/min (promo, was $0.0077) · Flux English $0.0065callback param$200Streaming diarization +$0.0020/min
Gladia Solaria-3€0.16/hr = €0.0027/min (€0.12 at 10,000+ hrs)Real-time €0.68/hr (€0.0113/min)callback_url + signed webhooks€50Solaria-3: 5 languages; training on by default (opt-out)

For accuracy rankings and open-weight options, see our best speech-to-text APIs ranking; for running Whisper-class models yourself, see best local speech-to-text models.

Batch transcription: price per minute

For files (meetings, calls, podcasts), the decision is price, language coverage and add-ons:

  • ElevenLabs Scribe v3 ($0.20/hour) — the cheapest multilingual list price; ElevenLabs says it makes 27% fewer errors than Scribe v2, supports 90+ languages, diarization of up to 32 speakers and number formatting.
  • AssemblyAI-STT-3.7 ($0.21/hour) — covers 32 languages; Universal-2 ($0.15/hour) handles the rest. Add-ons stack: Medical Mode +$0.15/hour. EU processing costs the same as US. Multichannel audio is billed per channel.
  • Speechmatics Melia 1 ($0.24/hour) — 55+ languages without choosing one up front and code-switching mid-sentence; annual credit plans cut the rate by up to 25%.
  • Deepgram Nova-3 ($0.0043/minute, $0.258/hour) — diarization and smart formatting included on pre-recorded audio; Growth plan ($4,000+/year prepaid) drops it to $0.0036.
  • Gladia Solaria-3 (€0.16/hour) — the cheapest rate here, but Solaria-3 covers 5 languages; Solaria-1 (€0.27/hour) covers 100+.

Streaming: watch the billing unit

Voice agents and live captions need streaming, and the billing unit matters as much as the rate. AssemblyAI bills the time the WebSocket is open, not the audio you send: a connection held for 60 minutes with 30 minutes of speech is billed 60 minutes. Deepgram bills per second of audio with no block rounding. At the time of writing:

  • ElevenLabs Scribe v3 Live — $0.22/hour (≈$0.0037/min), ElevenLabs claims 43% fewer errors than v2 Realtime and better turn detection.
  • Deepgram Nova-3 streaming — $0.0048/min promotional (list $0.0077); Flux English, built for voice agents with turn detection, $0.0065.
  • AssemblyAI-STT-3.7-Realtime — $0.30/hour for captions and notetakers; the Realtime-Agent variant for voice agents is $0.45/hour.
  • Speechmatics Linden 1 — $0.30/hour (introductory, from $0.40), with built-in end-of-turn signals.
  • Gladia real-time — €0.68/hour, the most expensive streaming rate in this table.

For the transport layer (WebRTC, SIP, phone numbers), see cloud providers for real-time audio streaming; for all-in voice agent costs, see voice agent platform pricing per minute.

Webhooks: how each API notifies you

  1. Deepgram: add callback=URL to /v1/listen; Deepgram returns a request_id immediately and posts the full result later. Authenticate with Basic Auth credentials in the URL or a secret in extra. Only ports 80, 443, 8080 and 8443 are allowed.
  2. AssemblyAI: set webhook_url on POST /v2/transcript; separate webhooks exist for streaming.
  3. ElevenLabs: set webhook=true; results go to the webhooks configured in your account, with webhook_metadata echoed back for correlation.
  4. Gladia: pass callback_url on POST /v2/pre-recorded, or configure account webhooks (transcription.created, .success, .error) delivered through Svix with signatures.
  5. Speechmatics: add a notification config to the batch job; allowlist the published egress IPs for your region (EU1, EU2, US1, US2, AU1).

Whichever you pick, store the job ID before the webhook arrives and make the handler idempotent: retries can deliver the same result twice.

Which to pick

  • Cheapest multilingual batch: ElevenLabs Scribe v3 or AssemblyAI-STT-3.7.
  • Cheapest if your languages are covered: Gladia Solaria-3.
  • Voice agents: Deepgram Flux or Nova-3 streaming, or AssemblyAI Realtime-Agent if you already use AssemblyAI.
  • Mixed-language audio: Speechmatics Melia 1.

Last verified: November 18, 2026. Promotional rates (Deepgram streaming, Speechmatics Linden 1) can end without notice; recheck before committing volume.

Sources