A practical reference for sorting speech-to-text APIs by the features that shape an integration: price, language coverage, streaming, and speaker diarization.
What a speech-to-text API is, who this directory is for, and how to read the board.
A speech-to-text API turns spoken audio into written text over a network call. This directory lists the providers that developers and technical buyers evaluate most often and puts them on the same terms: price, language coverage, real-time streaming, and speaker diarization, with a primary source behind every figure.
The goal is a fast, honest first pass. Each field is verified against the provider’s own pricing and documentation, native billing units and tiers are footnoted, and the fields that cannot yet be measured on equal footing, accuracy and latency, are left empty rather than filled with numbers that do not compare.
Use the board to narrow a shortlist, open a provider profile for the detail and the primary-source links, then confirm anything load-bearing in the provider’s own docs before you build. New to the terms? The speech-to-text glossary defines the vocabulary the field uses.
Use this as a starting point, not a final verdict. The current order is a transparent working sort based on streaming and diarization support, then listed price where available.
| # | Provider | Listed price | Languages | Streaming | Diarization | Free tier |
|---|---|---|---|---|---|---|
| 01 | Speechmatics | $0.0022/min | 55+ | Yes | Yes | Yes |
| 02 | AssemblyAI | $0.0035/min | 99+ | Yes | Yes | Yes |
| 03 | ElevenLabs Scribe | $0.0037/min | 90+ | Yes | Yes | Yes |
| 04 | Deepgram | $0.0043/min | 45+ | Yes | Yes | Yes |
| 05 | OpenAI (Whisper / gpt-4o-transcribe) | $0.0060/min | 98+ | Yes | Yes | No |
| 06 | Telnyx | $0.0150/min | 100+ | Yes | Yes | No |
| 07 | Google Cloud Speech-to-Text | $0.0160/min | 125+ | Yes | Yes | Yes |
| 08 | Azure AI Speech (Speech-to-Text) | $0.0167/min | 140+ | Yes | Yes | Yes |
| 09 | Amazon Transcribe | $0.0240/min | 100+ | Yes | Yes | Yes |
| 10 | Rev AI | $0.0033/min | 58+ | Yes | No | Yes |
| 11 | Google Gemini (audio understanding) | Varies | Not listed | Yes | No | Yes |
| 12 | Whisper (open source) | Self-hosted | 99+ | No | No | Yes |
Data status Pricing, languages, streaming, diarization, and free-tier fields verified against each vendor's primary pricing/docs pages on 2026-08-27 (see price_note, capability notes, and sources per provider). Prices are the cheapest documented async/pre-recorded per-minute-equivalent for the current flagship model; native billing units differ (per-hour, per-second, per-token) and are stated in price_note. Cheaper batch/volume tiers exist for several vendors and are described qualitatively where the exact rate is not publicly enumerated. wer_pct and latency remain unmeasured estimates and are NOT yet verified.
Each profile keeps the essentials in one place, with a direct route to primary docs when it is time to validate a fit.
Accuracy-focused STT with strong accent and dialect coverage.
Speech AI models plus audio-intelligence features on top of transcription.
ElevenLabs' speech-to-text model, the STT counterpart to its TTS lineup.
Purpose-built STT models with low-latency streaming and per-minute pricing.
Whisper-family transcription via the OpenAI API, plus newer gpt-4o transcribe models.
Transcription on Telnyx's own global network, co-located with voice and messaging APIs.
Hyperscaler STT with the widest language coverage and Chirp foundation models.
Microsoft's enterprise STT with custom-model training and broad locale support.
AWS-native transcription with tiered per-second pricing.
STT from a transcription company; human-accuracy transcription is a separate product.
LLM-native audio transcription via the Gemini API, priced on tokens not minutes.
Self-hosted open-weight model; no per-minute fee, you run the compute.
We show what is known, mark what is not, and will only add accuracy and latency comparisons once each provider is measured or sourced on the same basis. That makes the board useful without pretending its draft fields are settled facts.
A short grounding layer for buyers and the systems that need a direct answer before they explore the directory.
A speech-to-text API, also called automatic speech recognition or ASR, converts spoken audio into written text through a network call. A developer sends recorded or live audio and receives a transcript, often with timestamps, confidence scores, and optional speaker labels.
Compare speech-to-text APIs on four integration signals first: real-time streaming, speaker diarization, language coverage, and a pricing model that fits your volume. Once a shortlist forms, validate accuracy and latency yourself on audio that matches your use case, because vendor-reported numbers often come from different test sets.
Real-time, or streaming, transcription returns text while audio is still arriving, which suits live calls and captions. Batch, or pre-recorded, transcription processes a completed file and returns the transcript afterward, usually at a lower price. Many providers charge more for streaming than for the same audio processed in batch.
Cloud speech-to-text pricing usually runs from roughly $0.002 to $0.025 per minute, but the headline number hides real differences. Providers bill in different units, per minute, per hour, per second, or per token, and batch tiers are often far cheaper than streaming. Each provider profile lists the cheapest documented pre-recorded rate with the native unit and tier footnoted.
Speaker diarization separates and labels each distinct speaker in an audio stream, so a transcript reads as separate speakers rather than one block of text. It matters for meetings, call analytics, and interviews, and less for single-speaker dictation. Some providers document only speaker-change detection, which flags a switch without labeling each speaker.
Yes. OpenAI Whisper is open source under an MIT license, so it can be self-hosted with no per-minute fee, though you pay for the compute to run it. Most hosted APIs also offer free access, which the directory marks as a perpetual allotment, a time-limited trial, or a one-time credit.
No single speech-to-text API is most accurate for every case, because accuracy depends on audio quality, accents, domain vocabulary, and background noise. Test two or three providers on audio that matches your workload and measure word error rate on the same files. This directory leaves accuracy fields empty until every provider can be measured on one benchmark.
Explore speech-to-text built alongside voice and messaging APIs.
Build on Telnyx