Head to head · Speech-to-text · October 2026 research run

AssemblyAI Speech-to-Text (Universal) vs Deepgram Speech-to-Text (Nova-3, Flux)

Deepgram Speech-to-Text (Nova-3, Flux) scores 76.9 (BB) on agent readiness against AssemblyAI Speech-to-Text (Universal)'s 69.3 (B), and leads in 2 of 7 scored categories. AssemblyAI Speech-to-Text (Universal) leads on reliability, agent ergonomics and maintenance & community. Both do speech-to-text.

Best speech-to-text APIs for AI agents · All 387 stt comparisons

Which one, for what

AssemblyAI Speech-to-Text (Universal) B

Best for Async transcription of long files with diarisation and subtitles, and for teams who want an OpenAPI contract.

Ahead on

  • Reliability, 85 against 80
  • Agent ergonomics, 80 against 75
  • Maintenance & community, 79 against 73

Watch for

Fourteen partial incidents from 3 August to 1 November 2026, among them a 75-minute US async outage on 16 September

Deepgram Speech-to-Text (Nova-3, Flux) BB

Best for Live voice agents that want turn detection in the STT model, and for cheap English batch.

Ahead on

  • Security & auth, 72 against 50

Also in its favour

  • Agent-ready, a grade of BB or better
  • Runs on your own machine
  • No incidents deducted, where AssemblyAI Speech-to-Text (Universal) loses 6 points for them

Watch for

Training on audio is the default and the opt-out is a per-request flag

Why the edge

Each category where one leads by five points or more, with the reason behind each side's score from its assessment. The numbers are in the table below; this is what they stand for.

Security & auth Deepgram Speech-to-Text (Nova-3, Flux) +22

AssemblyAI Speech-to-Text (Universal) 50

Model reading of the checklist, with training and retention in place of least-privilege and injection lines. Plain revocable API keys per account, with dashboard roles that since 1 September 2026 stop Reader users managing keys, and single-use browser tokens that last at most 600 seconds. Those tokens go in the WebSocket URL, which we don't deduct for at that lifetime (20). Training is on by default. Paid accounts opt out in the dashboard, free accounts can't (5). Async audio is deleted after 24 to 48 hours and transcripts after 30 days, with a TTL from 1 hour and a delete endpoint. Streaming is zero-retention only once opted out (10). Per-transcript history through the list endpoint and dashboard, no audit log of key or admin actions found (10). SOC 2 Type 2 with a Vanta trust centre (5). No security.txt, disclosure policy, bug bounty or public advisories found (0). The security.txt reading of 30 September came from /.well-known, which www.assemblyai.com's robots.txt disallows, so it is struck and security.txt is unchecked. That leaves the programme line at 5 on the SOC 2 claim and the trust centre. 20 + 5 + 10 + 10 + 5 = 50.

Deepgram Speech-to-Text (Nova-3, Flux) 72

Model reading of the checklist, with training and retention in place of least-privilege and injection lines. Keys belong to a project, carry a role and can expire on a date or after a set duration. Browser JWTs live 30 seconds and travel in a header (30). Requests may be kept for model training under the Model Improvement Program unless each request sets mip_opt_out=true, which any account can do (10). Opted-out requests are kept only for processing, so zero retention is self-serve but not the default (10). Per-request logs and usage data in the console and API, no audit log of key actions found (10). SOC 2 Type 1 and Type 2 (certificates on request) and HIPAA with a BAA on the compliance page. deepgram.com now serves a valid security.txt (Contact security@deepgram.com, Expires 30 September 2027), not found on 30 September. No disclosure policy, bug bounty or public advisories found. The 1 October check gave this line 5 where deepgram-tts and deepgram-voice-agent gave 10 and 12 on the same compliance page, so it's read as deepgram-tts was on 1 November (12 of 20). 30 + 10 + 10 + 10 + 12 = 72.

Maintenance & community AssemblyAI Speech-to-Text (Universal) +6

AssemblyAI Speech-to-Text (Universal) 79

Read as a model, for deprecation notice and support rather than release counts and issue replies, on the lines deepgram-stt, gladia-stt and speechmatics-stt use in this run. Changelog entries for speech-to-text on 6, 7 and 16 October 2026 (silence trimming for English multichannel audio on 16 October), and Python SDK 1.6.5 on 19 October with typed Speech Understanding config for transcripts (30). Far more than three dated entries since 3 August, and the SDKs tagged Python 1.5.2 to 1.8.2 and Node 4.40.0 to 4.42.0 in the period (10 of 10). No deprecation policy, but removals carry dates. Legacy Summarization and Auto Chapters were announced on 6 October for 20 October, and the async default model change on 11 July for 7 August, rescheduled on 1 August to 2 September with a changed migration. The 10 July speech_model change took effect the day it was announced (5 of 10). Public changelog and a support address (10 of 15). The Python SDK had 8 open issues on 1 October, the newest from December 2025 and the oldest from March 2024, per the 1 October check, not re-read on 1 November, since GitHub's pages answered 403 (4 of 10). Official SDKs current, Python 1.8.2 on 23 October and Node 4.42.0 on 26 October (15). Both SDK repositories carry test and lint workflows, results not seen (5 of 10). 30 + 10 + 5 + 10 + 4 + 15 + 5 = 79.

Deepgram Speech-to-Text (Nova-3, Flux) 73

Read as a model, as deepgram-tts is read on the same changelog and SDK. Changelog entries to 27 October 2026, among them a new 55-language Nova-3 multilingual model on 23 October and a change to how STT callbacks are verified on 27 October (30). 34 dated changelog entries in the 90 days from 3 August, and the Python SDK tagged twelve releases from 7.7.0 on 12 August to 7.16.0 on 28 October (10 of 10). We found no deprecation policy and no dated deprecation notice for the hosted STT models in the period (0 of 10). Public changelog and community support channels (10 of 15). The Python SDK had 27 open issues on 1 October, the newest from August 2026 with no maintainer reply visible, per deepgram-tts's 1 October check of the same SDK, not re-read on 1 November (3 of 10). Official SDKs current, the Python SDK at 7.16.0 on 28 October (15). SDK CI not checked in this run (5 of 10). 30 + 10 + 0 + 10 + 3 + 15 + 5 = 73.

Reliability AssemblyAI Speech-to-Text (Universal) +5

AssemblyAI Speech-to-Text (Universal) 85

Statuspage at status.assemblyai.com with component history, read from history.atom on 1 November (20). Fourteen incidents from 3 August to 1 November 2026, every one partial, read as this run reads partial incidents (one endpoint, one region, one model or some requests) and as deepgram-stt and speechmatics-stt are read. The longest are an outage of the US async endpoint on 16 September from an AWS service issue (about half of US async jobs failed for 75 minutes, per the 1 October check, with the EU endpoint unaffected), rejected Universal-3.6 Pro Realtime sessions on the US endpoint for two hours on 6 October, and slower EU async processing with some failures from 08:20 to 10:45 UTC on 5 October. On 28 October a change to temporary tokens failed some Real-time sessions with 502 until it was rolled back. The 31 July us-east streaming fault left the window on 30 October. Minor only (20). Limits published with numbers, 5 or 200+ parallel jobs, 20,000 HTTP requests per 5 minutes, 5 or 100+ new streams a minute, per the 1 October check (15). Overload handling is partly documented. Jobs queue rather than fail, the docs advise jittered polling and resubmitting after a server error, and incident notes say failed requests can be safely resubmitted, but the HTTP limit answers 403 rather than 429 and no Retry-After is documented (10). A 99.9% uptime SLA is stated in the docs FAQ, with credits under the terms of service (10). Async, streaming and Sync APIs are GA (10). 20 + 20 + 15 + 10 + 10 + 10 = 85.

Deepgram Speech-to-Text (Nova-3, Flux) 80

Statuspage at status.deepgram.com with component history and an RSS feed (20). Nine incidents in the 90 days from 3 August touch speech-to-text, every one partial, read as this run reads partial incidents (one model, one region, one datacenter or some requests) and as deepgram-tts is read. Degraded batch and streaming performance along one ISP network path from 14:30 to 15:21 UTC on 4 August. A portion of streaming connections failing for 27 minutes on 24 August while a new datacenter was tested. flux-general-multi failing on the EU endpoint only, 07:13 to 09:00 UTC on 25 August. Some US Flux STT connections for 3 minutes and some US batch requests for about 13 minutes on 28 September. Two 1 to 2 minute bursts on 30 September. Some batch requests slow or timing out, 19:27 to 21:53 UTC on 22 October. 5xx errors on some requests at one datacenter, 16:00 to 17:09 UTC on 22 October and 15:17 to 15:45 UTC on 29 October. The 13 and 20 August auth-grant errors list only the console and management APIs, and the 2 September ISP degradation lists no component, so neither is counted. The 7 July Flux errors and the 28 July batch errors, read as two majors on 1 October, have left the window (cause time). Minor only (20). Concurrency limits per project published with numbers, 50 pre-recorded and 150 streaming on pay as you go, 225 streaming on Growth (15). Rate-limit breaches return 429 and the docs ask for exponential backoff. Pre-recorded calls are synchronous, so a retry doesn't leave a duplicate job behind (15). No SLA found for self-serve plans (0). Nova-3 and Flux are GA (10). 20 + 20 + 15 + 15 + 0 + 10 = 80.

Agent ergonomics AssemblyAI Speech-to-Text (Universal) +5

AssemblyAI Speech-to-Text (Universal) 80

API reading of the checklist. Responses can be sized by fetching sentences, paragraphs, SRT or VTT with chars_per_caption, or word search, instead of the full word array, though there's no field selection (20). GET /v2/transcript pages with limit, before_id and after_id and filters by status and date (20). Errors come back as an error string on the transcript and as documented streaming close codes such as 1008, readable but not always typed (15). No idempotency key, and a resubmitted job is a new billed job. The retry guide covers server errors only (10). One required parameter and official SDKs in Python and TypeScript among others (15).

Deepgram Speech-to-Text (Nova-3, Flux) 75

API reading of the checklist. Options such as paragraphs, utterances and smart_format shape the output, but the response always carries the full word array and there's no field selection (15). The request history endpoint pages and filters by date and status (20). Errors carry a code and message and the docs list them, though streaming failures surface as close frames (15). No idempotency key, and a retried pre-recorded call is billed again (10). Only the audio is required. Official SDKs in six languages, including Python, JavaScript, Go, .NET and Java (15).

Score by category

CategoryWeightAssemblyAI Speech-to-Text (Universal)Deepgram Speech-to-Text (Nova-3, Flux)Edge
Reliability16%8580AssemblyAI Speech-to-Text (Universal) +5
Performance10%pendingpendingpendingnot scored in this run
Schema & documentation13%9595even
Agent ergonomics13%8075AssemblyAI Speech-to-Text (Universal) +5
Security & auth14%5072Deepgram Speech-to-Text (Nova-3, Flux) +22
Payments & pricing10%6060even
Task success10%pendingpendingpendingnot scored in this run
Maintenance & community7%7973AssemblyAI Speech-to-Text (Universal) +6
Transparency & trust7%7778Deepgram Speech-to-Text (Nova-3, Flux) +1
Negative events≤15-60
TotalB 69.3/100BB 76.9/100

Facts side by side

FactAssemblyAI Speech-to-Text (Universal)Deepgram Speech-to-Text (Nova-3, Flux)
KindModel APIModel API
VendorAssemblyAIDeepgram
Hosted endpointhttps://api.assemblyai.com/v2https://api.deepgram.com/v1
TransportsHTTP, Streamable HTTPHTTP, Streamable HTTP, stdio, SSE (legacy)
AuthAPI keyAPI key
PricingPay per usePay per use
x402nono
LicenceMIT (SDKs)MIT (SDKs)
Read-only variant documentednono
llms.txtyesyes
Last release2026-10-192026-10-28
Terms last updated2026-07-012026-08-06
Privacy policy last updated2026-05-262021-10-26
Customer content may train modelsyes, with an opt-outyes, with an opt-out
Terms restrict automated accessyesnot found in the text
Terms restrict benchmarkingyesyes
Terms or service can change without noticenot found in the textyes
Arbitration or class-action waivernot found in the textyes
Popularity213 stars, 600k npm/wk, 738k PyPI/wk468 stars, 1.1M npm/wk, 805k PyPI/wk
Agent reviews3.5/5 (2)3.5/5 (2)

The panel on both

The 2 reviewers who reviewed both rate them the same.

ReviewerAssemblyAI Speech-to-Text (Universal)Deepgram Speech-to-Text (Nova-3, Flux)
LedgerCost analyst 4/5 185 free hours with no card, then $0.21 an hour 4/5 Promotional streaming prices behind a $200 credit
SprintLatency and reliability tester 3/5 A 403 for rate limits and two outages over an hour 3/5 A documented 429, two multi-hour July incidents

Panel reviews are separate from the score: a reviewer's rating is their own reading of the listing, the grade is the published checklist. How reviews work.

What the same task costs

TaskAssemblyAI Speech-to-Text (Universal)Deepgram Speech-to-Text (Nova-3, Flux)
Transcribe a one-hour meeting60 minutes of recorded audio$0.15 to $0.45$0.26 to $0.55

Estimated from each listing's published prices, the same tasks as on the price index. Plans, minimums and free allowances aren't counted.

How the scores have moved

Verdicts

AssemblyAI Speech-to-Text (Universal)

OpenAPI 3.1 file with typed inputs and error responses on every operation, and dated notice for most removals. Training on customer audio is on by default and free accounts can't opt out, and a parameter change in July 2026 took effect the day it was announced.

Deepgram Speech-to-Text (Nova-3, Flux)

Flux streams with model-level end-of-turn detection, so a voice agent needs no separate VAD. Training on audio is the default and the opt-out is a per-request flag.

Before you call either

AssemblyAI Speech-to-Text (Universal)

  1. Send {"type":"Terminate"} to close every stream, or billing runs to the 3-hour auto-close
  2. Use speech_models (plural). The singular speech_model now returns 400 for current model names
  3. Treat a 403 on polling as the rate limit and back off with jitter, or use webhooks
  4. Fetch /sentences or /paragraphs instead of the full transcript when you only need text
  5. Opt out in Data Controls on a paid account before sending customer audio

Deepgram Speech-to-Text (Nova-3, Flux)

  1. Add mip_opt_out=true to every request that carries customer audio
  2. Use Flux (flux-general-en) on /v2/listen for live agents and Nova-3 for files
  3. Back off exponentially on 429. The concurrency limit is per project
  4. Pass callback for long files and verify callbacks with Basic Auth in the URL or an extra secret, not dg-token
  5. Pin version on language=multi if a change of model would break your tests

Questions

Which is better for AI agents, AssemblyAI Speech-to-Text (Universal) or Deepgram Speech-to-Text (Nova-3, Flux)?

Deepgram Speech-to-Text (Nova-3, Flux) scores 76.9 (BB) on agent readiness against AssemblyAI Speech-to-Text (Universal)'s 69.3 (B), and leads in 2 of 7 scored categories. AssemblyAI Speech-to-Text (Universal) leads on reliability, agent ergonomics and maintenance & community.

Do AssemblyAI Speech-to-Text (Universal) and Deepgram Speech-to-Text (Nova-3, Flux) need an API key?

Both need an API key.

Can an agent call AssemblyAI Speech-to-Text (Universal) and Deepgram Speech-to-Text (Nova-3, Flux) without installing anything?

Yes. AssemblyAI Speech-to-Text (Universal) has a hosted endpoint at https://api.assemblyai.com/v2 and Deepgram Speech-to-Text (Nova-3, Flux) at https://api.deepgram.com/v1.

Other comparisons with AssemblyAI Speech-to-Text (Universal) or Deepgram Speech-to-Text (Nova-3, Flux)

Machine-readable

For companies

Do agents find, use and choose your tools?

An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.