Best Speech-to-Text APIs in 2026

Speech-to-text is the first link in every voice agent, and it fixes two things you cannot patch later: how soon your agent can start answering, and how many words are already wrong before the language model sees them. List prices run from $0.04 per hour to more than $0.90 per hour, and the accuracy ranking changes depending on whose benchmark you trust.

The short answer: on the independent Artificial Analysis table, ElevenLabs Scribe v3 (1.9% WER) and Microsoft MAI-Transcribe-2 (2.0%) lead the vendors covered here. For live agents, Deepgram Flux and AssemblyAI Universal-3.6 Pro Realtime are built around turn-taking. For the cheapest hosted option, Groq runs Whisper at $0.04 per hour, and for self-hosting, NVIDIA's Parakeet and Canary-Qwen are the open models to try first. If you are choosing a whole stack rather than one component, read the voice agent platforms comparison and pair this with the text-to-speech APIs guide.

Deepgram's own launch video for Flux, the first model in this guide that merges transcription and turn detection.

Quick Picks

These are the shortlists we would start from. Each line names the criterion, the pick, and the evidence behind it. All prices and scores are as of Nov 11, 2026.

Master Comparison Table

One row per product, with hourly prices converted from each vendor's published unit (per minute, per hour, or per 1,000 minutes). Where a vendor did not publish a number on the page we fetched, the cell says so instead of guessing. Gladia prices in euros and is left in euros.

Reading notes: AWS figures come from the worked examples on AWS's own pricing page (US East, tiered by volume), so treat them as illustrative and check your region. Google's $0.016 per minute is the first volume tier in the V2 table; the same table drops to $0.004 per minute above 2 million minutes a month. Hourly figures are our arithmetic from the per-minute prices.

Deepgram (Nova-3, Flux)

Deepgram sells two lines. Nova-3 is the general transcription model; Flux is a newer “conversational speech recognition” model that returns turn-complete transcripts with StartOfTurn and EndOfTurn events, so you do not stitch together a VAD, an endpointing timer, and an ASR stream. Deepgram says Flux matches Nova-3 on word error rate while cutting agent response latency by 200 to 600 ms against pipeline approaches and false interruptions by about 30% (vendor-reported, from its launch post).

Pricing (pay as you go): Flux English and Nova-3 monolingual streaming are $0.0077 per minute (about $0.46 per hour); Nova-3 multilingual streaming is $0.0092; Flux Multilingual is $0.0078. Pre-recorded Nova-3 is $0.0043 per minute monolingual and $0.0052 multilingual, and a hosted Whisper Large is $0.0048. Growth plans start at $4,000 a year in prepaid credits and lower these rates (for example $0.0065 per minute streaming Nova-3). New accounts get a $200 credit with no card. Diarization is included on pre-recorded audio and costs $0.0020 per minute on streams; keyterm prompting is $0.0013 per minute.

Strengths. The best-documented turn-detection controls in this guide: eot_threshold defaults to 0.7, eot_silence_threshold_ms to 5000, and an optional eager_eot_threshold fires EagerEndOfTurn 150 to 250 ms before the final event at the cost of 50 to 70% more LLM calls. Nova-3 has the second-highest speed factor in the independent table (545.6x). Flux Multilingual covers 10 languages (English, Spanish, German, French, Hindi, Russian, Portuguese, Japanese, Italian, Dutch) with code-switching.

Weaknesses. Nova-3 scores 5.2% on the independent AA-WER table, behind ElevenLabs, Microsoft, AssemblyAI, Gladia, and OpenAI's newest model. Flux is streaming-only, and the original Flux release was English-only with self-hosting still on the roadmap. Best for: voice agents where turn timing matters more than the last point of WER, and batch workloads that value throughput.

AssemblyAI (Universal)

AssemblyAI has split its catalog into batch and streaming rate cards. Universal-3.5 Pro is $0.21 per hour and Universal-2 is $0.15 per hour for pre-recorded audio. For streaming, Universal-3.6 Pro Realtime is $0.45 per hour base, while Universal-Streaming (English and multilingual) is $0.15 per hour. Its own pricing reference tells builders of voice agents to use Universal-3.6 Pro Realtime rather than the cheaper streaming tier, which is a vendor recommendation worth testing rather than taking on trust.

Billing details that change your invoice. Streaming is billed by session duration, meaning the time the WebSocket is open, not the audio you send, so idle connections cost money. Add-ons stack: speaker diarization is +$0.02 per hour in batch (standard), +$0.065 (experimental), and +$0.12 per hour in streaming; keyterms prompting is +$0.05 per hour on Universal-3.5 Pro and included on Universal-3.6 Pro Realtime; translation is +$0.06 per hour. The EU region costs the same as the US. New accounts receive $50 in credits.

Strengths. Universal-3.5 Pro scores 3.0% on AA-WER at $0.21 per hour list (Artificial Analysis estimates $3.50 per 1,000 minutes, the same figure), with a 100.5x speed factor. The add-on menu is the broadest here: PII redaction, entity detection, sentiment, and an LLM Gateway for summaries. Weaknesses. Per-feature billing makes the sticker price a floor, not a quote, and a few older features (auto chapters, summarization) are marked deprecated on the Universal-3.5 Pro path. Prompting-style add-ons list English, Spanish, German, and French. Best for: teams that want accuracy, an audio-intelligence toolbox, and a clear batch versus streaming price split.

OpenAI (gpt-transcribe, Whisper)

OpenAI's transcription lineup on its pricing page, as of Nov 11, 2026, has grown beyond Whisper. gpt-transcribe is $0.0045 per minute ($0.27 per hour). gpt-4o-transcribe is $0.006 per minute, with token rates of $2.50 input and $10.00 output per million; gpt-4o-mini-transcribe is $0.003 per minute; gpt-4o-transcribe-diarize adds speaker labels at $0.006 per minute; classic Whisper stays at $0.006 per minute. For live audio, gpt-live-transcribe and gpt-realtime-whisper are listed at $0.017 per minute, about $1.02 per hour.

Strengths. Simple per-minute pricing, a diarization-capable model, and a familiar SDK. GPT Transcribe scores 3.3% on AA-WER, ahead of GPT-4o Transcribe (4.0%) and GPT-4o Mini Transcribe (4.5%). Weaknesses. Live transcription is the most expensive streaming rate among the major APIs we priced, and the independent table shows slower speed factors (42.7x for GPT Transcribe, 38.2x for GPT-4o Transcribe) than Deepgram, Groq, or Parakeet hosts. AA also notes it chunks long audio for the GPT-4o transcribe models because of time limits. Best for: teams already on OpenAI who want one vendor and one bill, and batch jobs that need diarization without another contract.

ElevenLabs Scribe

ElevenLabs is best known for voice synthesis, but Scribe is now one of the top transcription products. Its pricing page lists Scribe v3 and Scribe v2 at $0.20 per hour and the realtime models (Scribe v3 Live, Scribe v2 Realtime) at $0.22 per hour, with keyterm prompting as an extra. The company announced Scribe v3 and Scribe v3 Live on Nov 10, 2026, a day before this article's date.

Strengths. The independent table gives Scribe v3 a 1.9% AA-WER at $3.33 per 1,000 minutes, the lowest among major commercial APIs and second overall to Alibaba's preview model; Scribe v2 follows at 2.2%. ElevenLabs says Scribe v3 has 27% fewer errors than v2 and the lowest diarization error it tested, and that Scribe v3 Live has 43% fewer errors than v2 Realtime (all vendor-reported). The pricing page lists 90+ languages. Scribe v2 Realtime is listed at about 150 ms latency (vendor-reported). Weaknesses. The speed factor is moderate (69.8x), the lineup is new enough that independent streaming data for Scribe v3 Live is thin, and the ecosystem of audio-intelligence add-ons is smaller than AssemblyAI's. Best for: accuracy-first transcription, multilingual content, and teams that already use ElevenLabs for voice. See the sibling text-to-speech guide for the other half of that pairing.

Speechmatics

Speechmatics publishes a list-rate table in dollars per hour. Pre-recorded: Melia 1 $0.24, Oak 1 $0.30, Standard $0.45, Enhanced $0.75. Real-time: Linden 1 $0.30 (a promotional rate, down from $0.40), Melia 1 $0.40, Standard $0.45, Oak 1 $0.46, Enhanced $0.80. The page says these exclude model-training and volume discounts, and new accounts start with $100 in credits.

Strengths. Melia 1 is a multilingual model that transcribes 56 languages without choosing one up front and switches languages mid-sentence; the vendor lists 55+ languages in total, plus language hints and code-switching. Linden 1 targets live voice agents and low-latency turns, and Oak 1 targets clinical conversations and dictation. Speaker diarization is available on the pre-recorded models and on most real-time models (the comparison table shows it unavailable for Melia 1 real-time). Weaknesses. On the independent table Enhanced scores 4.0% but is priced at an estimated $12.50 per 1,000 minutes, the highest of the specialist vendors; Melia scores 4.9% at $4.00, and Standard 5.1% at $7.50. Best for: mixed-language audio, accented speech, and enterprise buyers who want a Master Services Agreement and sector-specific models.

Gladia (Solaria)

Gladia is a Paris-based API vendor with prepaid-credit pricing in euros. Async pricing starts at EUR 0.16 per hour on Solaria-3 and EUR 0.27 on Solaria-1, falling with volume to EUR 0.12 and EUR 0.21 above 10,000 hours a month. Real-time is EUR 0.68 per hour. New accounts get EUR 50 in credits (valid one year), which Gladia says is roughly 312 hours of async transcription on Solaria-3.

Strengths. Solaria-3 scores 3.2% on AA-WER at an estimated $2.98 per 1,000 minutes, among the best accuracy-per-price combinations in the table, and Solaria-1 covers 100+ languages. Add-ons are cheap (summarization EUR 0.03 per hour, translation EUR 0.05). Weaknesses. Solaria-3 supports only 5 languages, diarization and enhanced punctuation are async only, real-time costs about four times the async rate, and self-serve accounts are opted in to model training by default (a free opt-out exists). Concurrency is 25 async and 30 real-time sessions on self-serve. Best for: cost-sensitive async transcription in a handful of languages, and European buyers who want a local vendor with GDPR, HIPAA, and SOC 2 Type II listed.

Google Cloud (Chirp)

Google's Speech-to-Text V2 prices recognition per minute with monthly tiers per account: $0.016 per minute for the first 500,000 minutes, $0.010 up to 1 million, $0.008 up to 2 million, and $0.004 above that. Dynamic batch, which processes audio at lower urgency, is $0.003 per minute. The V1 API includes 60 free minutes a month and charges $0.016 per minute with data logging or $0.024 without. The “Standard” price group includes the chirp model in V2 alongside default, latest_long, latest_short, and phone_call.

Strengths. Deep Google Cloud integration, volume tiers that fall steeply, and a cheap batch option. On the independent table, Chirp 3 scores 4.3% at $16.00 per 1,000 minutes; Google's Gemini 3.5 Transcribe, a separate LLM-based product, scores 2.6% at $5.00 and a 93.9x speed factor. Weaknesses. The older Chirp scores 31.2% in the table, so model naming matters, and Chirp 3 is slow (28.2x) and the most expensive per 1,000 minutes of any mainstream option listed. Best for: teams already on Google Cloud, high-volume batch (where the tiers pay off), and anyone who wants to try Gemini for transcription next to a classic ASR model.

Azure Speech

Azure's pricing page lists Real-time Transcription, Fast Transcription, MAI Transcribe-2, MAI Transcribe-1.5, Batch Transcription, LLM Speech Transcription, Custom Transcription, and add-ons for continuous language identification, diarization, and pronunciation assessment. The pricing page did not render dollar amounts when we fetched it, so we read the rates from Microsoft's public Azure Retail Prices API (East US, USD, queried Nov 12, 2026): real-time transcription $1.00/hr, Fast Transcription $0.36/hr, batch transcription $0.18/hr, and the enhanced-features add-on $0.30/hr. A Fast Transcription promo meter at $0.10/hr has been effective since Sep 1, 2026; we do not know its terms, so we do not treat it as the list price. We found no separate MAI-Transcribe-2 meter in that query. The page confirms a free F0 tier with 5 audio hours of real-time transcription per month, shared between standard and custom models, and that batch is not supported on the free tier. Rates vary by region, so check yours.

Independent data. Microsoft AI's MAI-Transcribe-2 scores 2.0% on AA-WER with a 311.8x speed factor at an estimated $1.67 per 1,000 minutes, the best combination of accuracy, speed, and price in the table for the models we cover. MAI-Transcribe-1.5 scores 2.4% at $6.00 and 77.3x. These are Artificial Analysis' normalized estimates, not Azure list prices. Weaknesses. Real-time at $1.00/hr is pricey next to the specialists, and the product catalog with many overlapping SKUs. Best for: Azure shops, and anyone who wants to test MAI-Transcribe-2 against Scribe v3 on their own audio.

AWS Transcribe

Amazon Transcribe bills per second with no minimum on standard transcription, in tiers that vary by region. AWS's own worked examples (US East, N. Virginia) price streaming at $0.01 per minute and batch at $0.006 per minute, with automatic content redaction at $0.0024 per minute in its first tier. The free tier is 60 audio minutes a month for the first 12 months. Call Analytics, Transcribe Medical, custom language models, and toxicity detection are priced separately.

Strengths. The integration surface: IAM, S3, Kinesis, and the rest of AWS, plus a two-channel conversation billed only for total duration. On the independent table Amazon Transcribe scores 4.1% at 39.1x and $6.00 per 1,000 minutes, and Amazon's newer Nova 2 Pro scores 4.9% at $3.10. Weaknesses. Mid-pack accuracy and speed against the specialists, and a tiered, region-dependent price card that is harder to compare than a flat per-hour rate. Best for: call-center and contact-center stacks built on AWS, where Call Analytics and data residency matter more than the last point of WER. If your agent also needs a place to run, see the GPU cloud guide.

Groq (hosted Whisper)

Groq hosts open Whisper models on its LPU inference stack. Its models page lists whisper-large-v3 at $0.111 per hour and whisper-large-v3-turbo at $0.04 per hour, with a 100 MB file-size limit and listed rate limits of 300 and 400 requests per minute respectively. Groq's docs describe large-v3 as the accuracy-oriented choice, supporting translation as well, and large-v3-turbo as a fine-tuned, pruned variant for fast multilingual transcription.

Strengths. The lowest list price in this guide, and the independent table shows Whisper Large v3 Turbo on Groq at 4.6% WER, a 312.8x speed factor, and $0.67 per 1,000 minutes. Weaknesses. It is batch transcription of an open model, so you get Whisper's limitations (repetition, uneven low-resource languages) and no built-in turn detection, diarization, or keyterm prompting that we could confirm. Groq's pricing page itself rendered empty, so our figures come from its documentation pages. Best for: cheap, fast batch transcription of recordings, prototypes, and pipelines where a post-processing LLM cleans up the text.

Open Models: Whisper, Parakeet, Canary-Qwen

Open weights remove the per-minute meter and put the cost into GPUs and operations. Three models matter in 2026, and they suit different jobs. Whichever you pick, benchmark it on the same audio as your API candidates, because Artificial Analysis shows the same weights scoring differently on different hosts.

Whisper large-v3

OpenAI's model card describes a 1550M-parameter multilingual model trained on 1 million hours of weakly labeled audio and 4 million hours of pseudo-labeled audio, with a 10% to 20% error reduction over large-v2 across languages. It is supported in Hugging Face Transformers, runs almost anywhere, and is the base for many hosted services (Deepgram, Groq, Together AI, fal.ai, and Replicate all list it). The card is candid about limits: uneven accuracy across languages, accents, and dialects, and a tendency to generate repetitive text. Independent scores range from 4.1% (fal.ai) to 10.1% (Replicate), which is a lesson in hosting variance.

NVIDIA Parakeet TDT 0.6B v3

A 600M-parameter multilingual model for high-throughput transcription, covering 25 European languages and released under CC BY 4.0. The model card reports an average WER of 6.34% on the Hugging Face Open ASR leaderboard datasets (AMI 11.31%, Earnings-22 11.42%, GigaSpeech 9.59%, LibriSpeech clean 1.93% and other 3.59%, SPGI 3.97%, TEDLIUM 2.75%, VoxPopuli 6.14%), and its multilingual tables list averages of 11.97% on FLEURS and 7.83% on MLS. The card includes a streaming inference script for NeMo. On Artificial Analysis, Together AI's hosted Parakeet TDT 0.6B V3 scores 4.5% at a 652.5x speed factor, the fastest in the table, for $1.50 per 1,000 minutes. Choose it for batch throughput on European languages.

NVIDIA Canary-Qwen-2.5B

An English-only speech-recognition and speech-LLM model with 2.5 billion parameters, released under CC BY 4.0. Its model card reports a 5.63 mean WER at 418 RTFx on the Hugging Face Open ASR leaderboard (AMI 10.18, GigaSpeech 9.41, LibriSpeech clean 1.60 and other 3.10, Earnings-22 10.42, SPGI 1.90, TEDLIUM 2.72, VoxPopuli 5.66; the card is dated July 2025, so treat these as a snapshot). Because it pairs the speech encoder with an LLM, it can also be prompted for summarization or question answering. Replicate's hosted version scores 4.3% on AA-WER but at a 7.9x speed factor, so deployment details change the result a lot. Choose it when you want English accuracy and LLM post-processing in one self-hosted model.

Benchmarks and Independent Data

We use three kinds of evidence: Artificial Analysis' speech-to-text leaderboard, fetched live on Nov 11, 2026; the Hugging Face Open ASR leaderboard, read on Nov 12, 2026; and the numbers NVIDIA prints on its model cards. Artificial Analysis' AA-WER v2 is an audio-duration-weighted average over about 8 hours: AA-AgentTalk (50%), VoxPopuli-Cleaned-AA (25%), and Earnings22-Cleaned-AA (25%). It uses cleaned reference transcripts, and where a model cannot handle full-length audio, AA chunks it (to about 9 minutes for some OpenAI and Amazon models, and about 30 seconds for models such as Canary Qwen). Price is AA's estimated cost for 1,000 minutes, normalized across billing models; speed factor is audio seconds transcribed per second of processing, median of the last 7 days on 10-minute files.

Source: Artificial Analysis speech-to-text leaderboard, non-streaming, fetched Nov 11, 2026. We show a subset (22 of the listed models) chosen to cover the vendors in this guide; the page also lists models from Soniox, Mistral, StepFun, Zoom, and others that we did not review. Scores change as models and hosts update, so treat the date as part of the number.

What the table says. The top of the list is tight: Scribe v3 (1.9%), MAI-Transcribe-2 (2.0%), and Voxtral Small Transcribe 3 (2.1%) differ by tenths of a point, which is within the range where dataset choice can reorder results. The middle (3% to 5%) holds most of the mainstream APIs, so price, speed, and features decide the pick. Deepgram's Nova-3 at 5.2% trades some accuracy for the table's second-best speed (545.6x). The old Chirp at 31.2% is a reminder to check which model an endpoint serves. One caveat: Artificial Analysis' own FAQ text lags its table (it still ranks Scribe v2 fourth), so we use the table values.

The Open ASR leaderboard, live. The Hugging Face Open ASR leaderboard (hf-audio/open_asr_leaderboard) renders in the browser, but its Space publishes the table data in a public Gradio config, which we read on Nov 12, 2026. The English tab had 64 ranked entries, the Multilingual tab 29 fully covered entries, and the Long-form tab ranked entries on Earnings21, Earnings22, and CORAAL. The table below keeps the models in this guide. Lower WER is better, and RTFx is throughput (audio seconds per second of compute) as the leaderboard reports it.

Source: Hugging Face Open ASR leaderboard Gradio config, fetched Nov 12, 2026. Three things stand out. First, the leaderboard now uses a newer dataset set (it includes AMI-Cleaned, Earnings22-Cleaned-AA-chunked, URGENT24, and private data), so its numbers no longer match the figures on NVIDIA's model cards: Canary-Qwen-2.5B shows a 6.92 average here against 5.63 on its card, and Parakeet TDT 0.6B v3 shows 7.04 against 6.34. Second, entries for Azure, ElevenLabs, AssemblyAI, OpenAI, and Gladia appear without an RTFx value, and Deepgram, Google, Amazon, and Groq do not appear in the English table we read, so a missing row does not mean a weak model. Third, we saw no ElevenLabs Scribe v3 entry, which is unsurprising two days after launch; the Scribe entry is v2. The leaderboard's ordering (Azure first on Multilingual, Scribe v2 first on Long-form) differs from the Artificial Analysis table, a reminder that dataset choice reorders results. For a retrieval layer to sit behind your transcripts, see best vector databases for AI agents.

Streaming Latency and Turn Detection

Latency numbers are the least comparable data in this market. Vendors measure different things (first partial word, end-of-turn, final transcript), on different audio, with different network paths, and there is no neutral streaming-latency table in our sources. What we can report, labeled by origin:

• Deepgram Flux: about 260 ms end-of-turn detection and 100+ streams per GPU (vendor-reported, launch press release); EndOfTurn fires within 1.5 s 95% of the time at default settings (vendor blog).
• ElevenLabs Scribe v2 Realtime: about 150 ms latency (vendor-reported, pricing page). Scribe v3 Live is listed with improved turn detection but no latency figure on the page we fetched.
• AssemblyAI Universal-3.6 Pro Realtime: positioned for voice agents; AssemblyAI claims it leads two independent streaming benchmarks, including a Pipecat open STT benchmark published by Daily (vendor-reported; we did not verify the benchmark pages).
• Speechmatics, Gladia, Google, Azure, AWS, Groq: no streaming latency figure on the pricing or docs pages we fetched. Groq and the other Whisper hosts serve batch requests.

For voice agents, raw milliseconds matter less than behavior at turn boundaries. A model that waits 700 ms to be sure you are finished feels slow; one that fires at 300 ms and cuts you off feels rude. Flux's eager end-of-turn setting and AssemblyAI's session billing both show how much the trade-off shapes cost: eager events can add 50 to 70% more LLM calls, and an idle socket is billed. If you are assembling the whole pipeline, compare LiveKit Agents vs Pipecat, and see how a CLI-driven voice workflow works in the Antigravity voice mode guide.

# Deepgram Flux (streaming voice agents), values from Deepgram's launch docs
wss://api.deepgram.com/v2/listen
  model=flux-general-en
  eot_threshold=0.7             # confidence needed to fire EndOfTurn (default 0.7)
  eot_silence_threshold_ms=5000 # silence fallback that forces a turn end (default 5000)
  eager_eot_threshold=0.3-0.5   # optional: EagerEndOfTurn arrives 150-250 ms earlier,
                                # at the cost of 50-70% more LLM calls
# events: StartOfTurn, EndOfTurn, EagerEndOfTurn, TurnResumed

# AssemblyAI streaming, per AssemblyAI's pricing reference
wss://streaming.assemblyai.com/v3/ws          # US
wss://streaming.eu.assemblyai.com/v3/ws       # EU (same price as US)
  speech_models: set explicitly, never rely on the default
  speaker_labels: true          # streaming diarization, +$0.12/hr
# billed for session duration (socket open time), so close the socket when the call ends

The block above is a reference sheet of parameters and endpoints taken from the vendors' launch notes and pricing references, not a runnable client; check each vendor's current API docs before you wire it up.

Diarization and Languages

Diarization (labeling who spoke) is packaged four ways. Included: Deepgram on pre-recorded audio. Add-on: AssemblyAI (+$0.02, +$0.065, or +$0.12 per hour depending on mode), Deepgram streaming ($0.0020 per minute), and Azure's enhanced-features add-on ($0.30/hr). Separate model: OpenAI's gpt-4o-transcribe-diarize. Async only: Gladia. Speechmatics lists diarization on its models, with the Melia 1 real-time exception noted above. ElevenLabs claims the lowest diarization error it tested for Scribe v3 (vendor-reported). There is no neutral diarization error-rate table in our sources, so test on multi-speaker audio from your own domain.

Languages are where vendor counts mislead. Gladia lists 100+ for Solaria-1 but 5 for Solaria-3. Deepgram lists 60+ for Nova but 10 for Flux Multilingual. ElevenLabs lists 90+ for Scribe. Speechmatics lists 55+, with Melia 1 transcribing 56 languages without a language hint. Parakeet TDT 0.6B v3 covers 25 European languages, Canary-Qwen is English only, and Whisper large-v3 is multilingual with uneven quality across languages. The Parakeet card shows how uneven: FLEURS WER of 3.00% for Italian against 24.03% for Slovenian. Count languages per model, per mode (batch or streaming), and test the ones you actually serve.

How to Choose

Start from the use case, shortlist two or three candidates, and run all of them on the same sample. The table below maps common scenarios to the options that fit best.

A practical test plan: collect 2 to 5 hours of real audio with its worst cases (phone quality, accents, crosstalk, domain jargon), write careful reference transcripts for a subset, and compute WER with one fixed normalizer for every vendor. For streaming, also record end-of-turn timing and false cutoffs on scripted interruptions. Then price the winner with add-ons included: diarization, redaction, keyterms, and any idle-connection billing. Teams running agents in the cloud may also want a sandbox for the code half of the stack; see browser infrastructure for AI agents and the best AI coding agents.

How We Evaluated

No paid placement in this ranking. There are no affiliate links, sponsored labels, or vendor reviews of the draft. Order within sections reflects the topic flow, not a score.

• Prices: each vendor's own pricing page, fetched Nov 11, 2026. Per-minute and per-1,000-minute prices were converted to per hour by multiplying; hourly values are our arithmetic.
• Accuracy and speed: Artificial Analysis AA-WER v2 and speed factor (non-streaming), fetched the same day, plus the Hugging Face Open ASR leaderboard (read Nov 12, 2026) and NVIDIA model-card figures.
• Vendor claims: labeled vendor-reported, including latency, error-reduction percentages, and benchmark claims in launch posts.
• Gaps we did not fill: an Azure MAI-Transcribe-2 list price (the page did not render; we used the Retail Prices API for the other meters), Google and AWS language lists, a neutral streaming-latency table, and Open ASR leaderboard rows for Deepgram, Google, Amazon, and Groq (none were listed). We say so rather than estimate.
• Not tested by us: we did not run these APIs on our own audio for this article. Rankings are a synthesis of published data.

FAQ

Sources

All pages fetched Nov 11, 2026 unless noted.

• Artificial Analysis: speech-to-text leaderboard (AA-WER v2, speed factor, price)
• Hugging Face: Open ASR leaderboard and its GitHub repository
• Deepgram: pricing
• Deepgram: Introducing Flux and its changelog entry
• Business Wire: Deepgram launches Flux (Oct 2, 2025)
• AssemblyAI: pricing
• OpenAI: API pricing (transcription)
• ElevenLabs: API pricing (Scribe)
• Speechmatics: pricing
• Gladia: pricing
• Google Cloud: Speech-to-Text pricing
• Microsoft Azure: Speech Services pricing
• AWS: Amazon Transcribe pricing
• Groq: speech-to-text docs and models
• OpenAI: whisper-large-v3 model card
• NVIDIA: parakeet-tdt-0.6b-v3 model card
• NVIDIA: canary-qwen-2.5b model card
• Deepgram on YouTube: Meet Flux (video embedded above)