Speechmatics Pricing in 2026: What a Voice Agent Pays per Minute

Speechmatics bills speech-to-text per second at hourly rates. Linden 1, its voice-agent model, costs $0.30 an hour, or half a cent per call minute. Melia 1 is $0.40 and Enhanced $0.80. Dograh, the open-source AI voice agent platform we build, runs Linden 1 on your own Speechmatics key. You can self-host it or run it on Dograh Cloud.

Key Takeaways

• Linden 1 at $0.30 an hour works out to $0.005 per call minute.
• The "from $0.12 an hour" headline is a discounted rate for recorded files.
• The opt-in training discount takes Linden 1 to $0.20 an hour.

Speechmatics lists its prices per hour of audio, but a voice agent gets billed one call at a time. Below, we turn each model into cost per call minute, go through the discounts and add-ons, set Linden 1 beside other streaming options, and finish with how it runs in Dograh.

How Speechmatics bills speech-to-text

Speechmatics charges by the second at an hourly rate, so a 47-second call is billed for 47 seconds. Speechmatics counts usage in credits, and one credit is $1 before discounts. Pay as you go bills the listed hourly rate with no discount and no commitment.

A new account gets $100 in credits with no card needed, and those signup credits last 12 months. On the free tier you can run 2 real-time sessions at once, and API access pauses when the credits run out.

Paying customers on the self-serve plan get 50 concurrent real-time sessions. For a voice agent, one session is one live call, so 50 is your ceiling on calls at the same moment. Enterprise plans have no fixed number, and Speechmatics says the limit adjusts to demand.

Every Speechmatics model, priced per minute

Each model has its own rate, and the one a voice agent should price first is Linden 1. Here is the full list, with each hourly rate turned into a per-minute figure by our arithmetic.

Linden 1 is the model behind Agent STT, Speechmatics' streaming endpoint for voice agents. It returns speaker-labelled segments with turn messages and runs only on that endpoint, away from the batch and regular real-time APIs. That is why it has no pre-recorded price.

The other models do different jobs. Melia 1 transcribes more than 55 languages without picking one first. Standard and Enhanced have the same features and differ in accuracy and speed, and Oak 1 is tuned for healthcare audio.

Here is a labelled example. One five-minute call on Linden 1 costs $0.30 / 60 x 5 = $0.025. The same call on Enhanced costs $0.067, more than two and a half times as much.

Discounts and the $0.12 headline

The "from $0.12 an hour" headline is a real price for recorded files, and a live call is billed well above it. Speechmatics shows how it is reached. Melia 1 on pre-recorded files is $0.24 an hour, the model-training discount takes 33% off, and the top annual plan takes another 25%. That leaves about $0.12, on a $15,000-a-year commitment.

The model-training discount is the one most voice teams can use. It is opt-in and off by default, and it takes 33% off speech-to-text rates. Opting in lets Speechmatics use your audio and transcripts to improve its models. If you stay out, Speechmatics says real-time inputs and outputs are never stored, which is why teams with health or payment details on their calls usually leave it off.

The other discounts depend on how you pay. Credit packs give up to 20% off and last 12 months. Monthly subscriptions give up to 15%, and annual ones up to 25%, starting at $120 a year for 5% off.

Our advice is to budget a live call at $0.30 an hour on Linden 1. Treat the training discount, which brings it to $0.20, as a separate decision about your callers' data, and treat the subscription discounts as a reward for volume you have already proven.

Open Source Alternative to Vapi / Retell

Self-hosted voice agent platform — no per-minute fees

dograh-hq/dograh

Star on GitHub

Add-ons, voices and Enterprise terms

Most add-ons are priced per hour on top of the transcription rate, and most voice agents need none of them. Translation is $0.65 an hour, chapters $0.40, topics $0.20, and summaries and sentiment $0.12 each. They suit recorded calls you review later more than a live conversation.

Speechmatics also sells text-to-speech at $0.011 per 1,000 characters, with the first million characters free. The training discount does not apply to it.

Security terms are the same on self-serve and Enterprise. Both cover SOC 2 Type II, ISO/IEC 27001:2022, GDPR and HIPAA, and you choose whether data is processed in the EU, the US or Australia. Single sign-on is an Enterprise add-on at $280 a month.

On-prem is an Enterprise option, with containers, a virtual appliance, on-device and hybrid setups. One limit matters for agents. Agent STT runs on Speechmatics' cloud only, so Linden 1 stays a hosted service even on an Enterprise contract. You can pin the region for data residency.

Speechmatics next to other speech-to-text APIs

On list price per minute, Linden 1 is one of the lowest rates among the streaming options a voice agent would test. Each rate below is the vendor's own list price, turned into a per-minute figure by our arithmetic.

The meters differ as much as the rates. Speechmatics and Deepgram bill per second of audio, as our breakdown of Deepgram's rates explains. AssemblyAI bills the time a session stays open, idle included, which our post on AssemblyAI pricing works through.

As a labelled example, 1,000 five-minute calls are 5,000 minutes, or 83.3 hours. That is $25.00 on Linden 1, $16.67 with the training opt-in, $37.50 on AssemblyAI's Universal-3.6 Pro and $38.50 on Deepgram Nova-3. Our Gladia pricing post walks through a whole call line by line.

Price is only half the test. On Pipecat's streaming STT benchmark, 1,000 conversational clips, linden-1 had a pooled word error rate of 0.96% with a median of 369 ms to its final segment. Deepgram's nova-3-general scored 1.37% at 247 ms, and AssemblyAI's universal-3-6-pro 0.80% at 307 ms. These clips are not phone calls, so test your own recordings before you choose.

Check the session cap too. Speechmatics' 50 self-serve sessions is a third of the 150 streams Deepgram allows on pay as you go. Our list of the best speech-to-text APIs compares those caps for each vendor.

Running Speechmatics in Dograh

Speechmatics is one of the speech-to-text providers built into Dograh, and it runs on your own key. You add Speechmatics under Providers in the sidebar, then choose it as the transcriber in a configuration under Models, as the Model Configurations guide shows. Speechmatics then bills you directly at its own rate.

The model that runs is Linden 1. Older saved setups that name Standard or Enhanced also run as Linden 1, so the rate that applies is $0.30 an hour, or $0.20 if your Speechmatics account opts in to training. Dograh's own turn-taking decides when the caller has finished, and Linden 1 supplies the words.

Words in the agent's Dictionary reach Speechmatics as additional vocabulary, which helps with product names and terms such as KYC. Our post on using a custom dictionary for industry jargon and names covers which words are worth adding.

Speech is one line of the bill. Self-hosted Dograh has no per-minute platform fee, but servers, model keys and phone minutes still cost money. Dograh Cloud charges a 1¢-a-minute platform fee. So a Cloud minute with Linden 1 is $0.005 from Speechmatics plus $0.01 of platform fee, before the other model bills and the phone line.

Each call's View Trace shows the speech-to-text time for every turn, and built-in QA tags miscommunication on each step of the call, so misheard words show up before you scale. If another provider does better on your recordings, the switch is a settings change, and the transcriber settings list every option, including self-hosted Whisper through Speaches.