productivity

Best AI for Transcription

The best AI transcription tools and speech-to-text APIs as of June 2026 — Fathom, Granola, Fireflies, Deepgram Nova-3, AssemblyAI, ElevenLabs Scribe v2, Mistral Voxtral and Whisper, with word error rates, pricing and recommendations for meetings, creators, developers and enterprise.

Updated June 2026

Quick answer: For meeting transcription, Fathom is the best free pick — unlimited recordings and transcripts at no cost — while Granola is the best bot-free option, recording the call on your own device with no bot joining, and Fireflies.ai leads for sales teams at $10/user/month with deep CRM sync. For developers, Deepgram Nova-3 (5.26% word error rate, from $0.26/hour) is the default for voice agents, while Mistral Voxtral Mini Transcribe 2 is the price-performance winner at $0.003/minute ($0.18/hour) and roughly 4% WER. On raw accuracy the field has plateaued: Deepgram Nova-3, AssemblyAI Universal-3 Pro, OpenAI gpt-4o-transcribe, ElevenLabs Scribe v2 and Microsoft MAI-Transcribe-1 all sit within about two percentage points of each other on clean English audio. For privacy or offline work, NVIDIA Parakeet-TDT-0.6B-v3 and open-weight Whisper run locally and keep audio on your machine.

One caveat before you record: AI meeting notetakers are now a live legal risk. Otter.ai faces a consolidated federal class action in the Northern District of California alleging its OtterPilot bot recorded Zoom, Microsoft Teams and Google Meet calls without participants’ consent, with legal analysts estimating a potential settlement in the $75–200 million range (NPR). Get consent before recording, and prefer bot-free or on-device tools for sensitive conversations.

This guide covers the full stack — consumer meeting assistants, creator tools, developer speech-to-text APIs, and open-source models you can self-host — with current word error rates, pricing and real user sentiment. Two things have changed the landscape since our last update: accuracy on clean English stopped being the differentiator (the top models cluster at 2–5% WER), and the competition moved to streaming latency, end-of-turn detection, code-switching and cost — while a wave of privacy litigation reshaped how the consumer tools are allowed to behave.

The current state of AI transcription: June 2026

AI transcription solved clean-English accuracy, and the market knows it. Independent voice-AI evaluator Coval calls the current state a “WER plateau”: the leading providers “sit within 1–2 percentage points of each other on LibriSpeech and FLEURS,” and “the competitive surface has shifted to streaming latency, end-of-turn detection, multilingual depth, code-switching, and cost at production scale” (Coval, June 2026). The market backing this is large and growing: the AI transcription market is projected to rise from about $4.5 billion in 2024 to $19.2 billion by 2034 (Market.us).

Five shifts define the current moment.

1. Word error rate plateaued — real-world conditions now decide. The top tier (Deepgram Nova-3, AssemblyAI Universal-3 Pro, OpenAI gpt-4o-transcribe, ElevenLabs Scribe v2, Microsoft MAI-Transcribe-1) all post 2–5% WER on clean English benchmarks. The differences that matter now show up on messy audio: alphanumeric IDs, proper nouns and code-switching. Coval notes that accuracy on order numbers and drug names “drops to 50–70% in many providers’ production output,” and a model at 5% WER on monolingual audio “routinely posts 15–20% WER on Spanish-English or Hindi-English code-switched calls.”

2. The privacy reckoning became real litigation. The Otter.ai class action that began in August 2025 is now a consolidated federal case in San Jose, with Otter moving to dismiss and the wider AI-notetaker industry effectively on trial over wiretap and consent law (UC Today). The practical effect: visible-bot, record-everything tools now carry compliance risk that on-device and bot-free tools avoid.

3. Bot-free notetaking won the consumer race. Granola — which captures system audio locally on a Mac with no bot in the meeting — raised $125 million at a $1.5 billion valuation in March 2026, up from a $250 million valuation less than a year earlier, with revenue up 250% in the prior quarter (TechCrunch). Bot-free is now a category, not a feature.

4. Speech-to-text split into “conversational” and “batch”. Voice agents drove a new model class with built-in turn-taking. Deepgram Flux Multilingual (April 2026) is the first conversational STT with integrated end-of-turn detection, removing the external voice-activity-detection layer and cutting 200–600ms off agent response time. OpenAI shipped a dedicated streaming model, GPT-Realtime-Whisper, in May 2026, separating low-latency STT from the batch Whisper line for the first time.

5. Open weights and aggressive pricing reset the floor. Mistral’s Voxtral Mini Transcribe 2 delivers roughly 4% WER on FLEURS at $0.003/minute — the best price-performance of any transcription API — and ships a sub-200ms open-weight realtime sibling under Apache 2.0 (Mistral). NVIDIA’s Parakeet-TDT-0.6B-v3 tops the open-source Hugging Face leaderboard, and hosted Whisper runs as low as $0.04/hour on Groq. Self-hosted, data-sovereign transcription is now genuinely good and genuinely cheap.

Top AI transcription models and APIs (June 2026)

Two cautions before the numbers. First, vendor word error rates are measured on each vendor’s preferred audio — Deepgram benchmarks a proprietary 2,703-file set; AssemblyAI publishes English-only numbers; OpenAI tends to compare against an older Whisper. They are not directly comparable. Second, streaming WER typically runs 1–3 points worse than batch WER on the same model. Treat vendor figures as a ceiling, treat the Hugging Face Open ASR Leaderboard as the floor for open models, and measure on your own audio before committing.

Developer speech-to-text APIs and models

Word error rates below are vendor- or benchmark-reported as noted; lower is better. Prices are list, per audio hour, before add-ons.

Two structural facts stand out. Mistral Voxtral undercuts the market by an order of magnitude — at $0.18/hour it is roughly a fifth of ElevenLabs Scribe v2’s cost while matching it on quality and running about 3x faster, per Mistral’s own benchmarks. And the headline-accuracy crown is genuinely contested: Microsoft MAI-Transcribe-1 claims the lowest FLEURS WER (3.8% across 25 languages), beating Whisper Large v3 on all 25, but it is a vendor number on a clean benchmark and Microsoft’s first proprietary speech model, launched April 2026 in Azure AI Foundry (Microsoft).

Independent and open-source benchmarks

For open models, the Hugging Face Open ASR Leaderboard is the standard — it ranks 86 systems across 12 datasets by word error rate and speed. As of mid-2026 the leaders combine a Conformer encoder with an LLM decoder: NVIDIA Canary-Qwen 2.5B tops the English track at roughly 5.6% average WER, alongside IBM Granite-Speech-3.3-8B and Microsoft Phi-4-Multimodal. For self-hosting with multilingual coverage, NVIDIA Parakeet-TDT-0.6B-v3 posts 6.34% average WER across 25 European languages and is the strongest open-source option you can run yourself (Coval).

Why benchmark scores disagree — and which to trust

A model can post 4% WER in a launch post and 15% on your call-centre audio. Both are real; the difference is the audio profile, not the model. Clean studio benchmarks (LibriSpeech, FLEURS) reward the demo; speakerphone calls with hold-music bleed, accents and overlapping speech reward whichever model your traffic happens to suit. The single most reliable predictor of production accuracy is WER measured on your own representative audio, scored not just on WER but on entity preservation (order numbers, drug names) and code-switching. Before paying a premium for a “more accurate” model, test two or three on a sample of your real recordings.

Best AI transcription for meetings

These tools capture business conversations and layer on summaries, action items and CRM sync. The 2026 dividing line is whether a visible bot joins your call.

Fathom — best free meeting transcription

Price: Free (unlimited recordings) | Premium $16/mo | Team $14/user/mo | Business $20/user/mo Platforms: Web, desktop, Zoom/Meet/Teams Bot: Yes (visible)

Fathom offers the most generous free tier in the category — unlimited recordings, transcripts and storage at no cost — and holds a 5.0/5 rating on G2 from over 6,000 reviews, the highest in the category. It claims around 95% transcription accuracy with 30-second post-call summaries. The paid tiers add unlimited AI summaries and team features.

Best for: Individuals and small teams who want unlimited meeting transcription without paying, and anyone testing AI notetakers before committing.

Granola — best bot-free notetaker

Price: Free tier | Pro and Business plans | Enterprise (with API) Platforms: macOS (primary), with expanding support Bot: None — records device audio locally

Granola is the breakout of 2026. It records system audio directly on your Mac, so no bot joins the meeting and no one on the call sees a recorder — then enhances your own typed notes with AI context. Its March 2026 round valued it at $1.5 billion, and it has expanded into team workspaces (Spaces) and a personal/enterprise API, with customers including Vanta, Gusto, Asana, Cursor, Lovable and Mistral AI (TechCrunch).

Best for: Solo professionals and Mac-first teams who want notes without a bot in the room, and privacy-sensitive conversations where a visible recorder is a non-starter.

Fireflies.ai — best for sales teams

Price: Free (800 min storage/seat, 20 AI credits/mo) | Pro $10/user/mo | Business $19/user/mo | Enterprise $39/user/mo (annual) Platforms: Web, Chrome, iOS, Android Bot: Yes (visible)

Fireflies.ai is the sales-team favourite, with support for 100+ languages, broad CRM coverage (Salesforce, HubSpot, Zoho) and “AskFred” natural-language search across your meeting history. Annual billing lands Pro at $10/user/month; monthly billing is $18. The free plan now offers unlimited transcription credits but only 800 minutes of storage per seat and a small one-time AI-credit pool.

Best for: Sales and revenue teams that need CRM auto-sync and a searchable meeting archive.

tl;dv — best for global teams

Price: Free (unlimited recording, auto-deletes after 3 months) | Pro ~$18/mo | Business $59/mo Platforms: Web, Chrome Bot: Yes (visible)

tl;dv matches Fathom’s unlimited free recording (though free transcripts auto-delete after three months) and adds 30+ languages plus the deepest integration ecosystem via Zapier. Strong fit for distributed teams that need multilingual notes and workflow automation.

Best for: Globally distributed teams and automation-heavy workflows.

Otter.ai — proceed with consent

Price: Free (300 min/mo, 30 min/conversation) | Pro $8.33/mo annual ($16.99 monthly), 1,200 min | Business $20/user/mo annual ($30 monthly) | Enterprise custom Platforms: Web, iOS, Android, Chrome Bot: Yes (OtterPilot, visible)

Otter.ai pioneered consumer transcription and still has excellent real-time display, a polished mobile app and a 20% student discount. But it is the defendant in the consolidated AI-notetaker class action over recording without consent (NPR), and its auto-join behaviour has drawn IT-level blocks. If you use it, confirm everyone on the call consents to recording.

Best for: Students and journalists who value real-time display — with explicit consent in place.

Krisp and Tactiq — noise and privacy specialists

Krisp pairs unlimited transcription with two-way noise cancellation and now offers VIVA voice isolation that sits in front of any speech-to-text engine, cutting WER 10–30% on noisy audio. Tactiq runs as a Chrome extension that displays a live transcript without a bot and without storing audio — a strong bot-free alternative to Granola for non-Mac users.

Best AI transcription for content creators

These tools fold transcription into a production workflow for podcasters, YouTubers and video editors.

Descript — best for text-based video editing

Price: Free | Hobbyist $16/mo annual ($24 monthly) | Creator $24/mo annual ($35 monthly) | Business $50/mo annual ($65 monthly) Platforms: Web, Mac, Windows

Descript still owns the paradigm: edit video by editing the transcript, delete words to cut footage, clean audio with Studio Sound, and patch mistakes with voice cloning. Pricing was restructured upward in 2025 and now starts at $16/month (Hobbyist, annual). The editing workflow is the draw; transcription accuracy trails dedicated APIs on hard audio.

Best for: Podcasters and YouTubers who edit their own content and value text-based editing over raw transcript accuracy.

Riverside and Sonix — recording and batch

Riverside.fm records broadcast-quality local audio and video for remote interviews and transcribes in 100+ languages as part of the workflow (from $15/month). Sonix offers transparent pay-as-you-go transcription at around $10/hour with professional Premiere, Final Cut and Audacity integrations and no subscription required — the simplest choice for irregular, high-quality batch jobs.

Rev — best when you need human accuracy

Rev remains the reference for high-stakes content. Its AI transcription runs about $0.25/minute at a claimed 96%+ accuracy, while human transcription at $1.99/minute guarantees 99%+ with professional formatting and 12–24 hour turnaround. For developers, Rev also ships the open-source Reverb model and hosted endpoints from $0.10–0.30/hour. Legal, medical and compliance work justify the human premium; everything else now has cheaper AI options.

Best speech-to-text APIs for developers

If you are building a product — captions, voice agents, call analytics — you want an API, not a desktop app. The current leaders:

Deepgram — best for voice agents at scale

Deepgram is the default for high-volume, low-latency products. Nova-3 delivers a vendor-reported 5.26% WER (batch) with sub-300ms streaming, from $0.26/hour batch. The April 2026 Flux Multilingual model adds integrated end-of-turn detection (median under 300ms), removing the external VAD layer that adds 200–600ms to agent turn-taking. Deepgram also claims 90%+ accuracy on alphanumeric IDs versus 43–58% across competitors — the entity-preservation edge that matters for order numbers and account IDs. A bundled Voice Agent API runs $0.075/minute.

Best for: Voice agents, real-time captioning and any product where turn-taking latency and entity accuracy are the bottleneck.

AssemblyAI — best for audio intelligence

AssemblyAI wins when you need transcription plus structured insight in one call. Universal-3 Pro (February 2026) introduced a “speech language model” with natural-language keyterm prompting (up to 1,500 words) to bias vocabulary, and posts 5.6% mean English WER at $0.21/hour async, $0.45/hour streaming. Built-in diarisation, sentiment, entity detection, PII redaction and summarisation make it the one-API choice for conversation analytics. Universal-2 (99 languages) remains for high-language-count workloads.

Best for: Teams that need ASR and NLU (sentiment, summaries, PII redaction) from a single API.

OpenAI — best for OpenAI-native pipelines

OpenAI’s speech-to-text now splits three ways: Whisper Large v3 / Turbo (offline batch, Turbo about 8x faster), gpt-4o-transcribe and gpt-4o-mini-transcribe (GPT-4o-class accuracy with streaming, at $0.006 and $0.003/minute), and GPT-Realtime-Whisper (May 2026, $0.017/minute streaming, paired with the Realtime-2 voice models). OpenAI reports gpt-4o-transcribe at 4.1% WER versus Whisper v3’s 5.3%. The trade-off: the realtime models lock you to OpenAI’s stack.

Best for: Products already in the OpenAI ecosystem, and native speech-to-speech agents.

ElevenLabs Scribe v2 and Mistral Voxtral — accuracy and value

ElevenLabs Scribe v2 is the batch model in the ElevenLabs voice stack, supporting 90+ languages with keyterm prompting and entity detection; the Scribe v2 Realtime sibling delivers sub-150ms latency, and a May 2026 pricing reset cut speech-to-text up to 45%. Mistral Voxtral Mini Transcribe 2 is the value leader at $0.003/minute with roughly 4% FLEURS WER, diarisation, context biasing and up to 3-hour files — and its Voxtral Realtime model is open-weights under Apache 2.0 for on-device deployment.

Enterprise clouds — Microsoft, Google, Amazon

Microsoft MAI-Transcribe-1 (April 2026) claims a category-leading 3.8% FLEURS WER across 25 languages at roughly half the GPU cost of alternatives, in Azure AI Foundry. Google’s Chirp 3 covers 125+ languages ($0.36/hour batch, $1/hour real-time), and Amazon Transcribe spans 100+ languages with HIPAA-eligible Medical and Call Analytics SKUs ($1.44/hour base). Pick by which cloud you already run in.

Best open-source and local transcription

For privacy, data sovereignty or high-volume cost control, self-hosting now matches commercial accuracy. The open frontier moved well beyond vanilla Whisper.

• NVIDIA Parakeet-TDT-0.6B-v3 — the strongest self-host option, 6.34% average WER across 25 European languages and top of the open-source leaderboard. GPU time only, no per-minute fee. Best for teams with ML infrastructure and high volume.
• Mistral Voxtral Realtime — open weights under Apache 2.0, sub-200ms latency, 4B parameters, edge-deployable in 13 languages (Mistral). The new open option for privacy-first realtime voice.
• faster-whisper — the practical general-purpose local engine, around 4x faster than vanilla Whisper at the same accuracy, with INT8 CPU support. MIT-licensed.
• whisper.cpp — best for edge and mobile, running on iPhone, Android, Raspberry Pi and even the browser via WebAssembly, with Apple Metal acceleration.
• WhisperX — adds accurate word-level timestamps and speaker diarisation, the complete open stack for meeting transcription.
• MacWhisper — the easiest local GUI for Mac, a one-time purchase that keeps everything on-device.

For teams that want Whisper economics without running GPUs, hosted Whisper is now extremely cheap: Groq at $0.04/hour, Lemonfox at $0.17/hour — far below OpenAI’s first-party Whisper endpoint.

Specialised and vertical tools

Some industries need domain accuracy and compliance that general tools can’t offer. The AI medical-transcription segment alone is forecast to keep growing sharply.

Medical. Nuance DAX (now Microsoft) leads ambient clinical documentation, generating EHR-ready notes after each visit. Amazon Transcribe Medical offers HIPAA-eligible API access with PHI identification, and Suki AI provides a mobile-first clinical assistant with EHR integration.

Legal. Verbit pairs ASR with human transcribers for near-100% accuracy on depositions and hearings; Rev and TranscribeMe offer court-grade human transcription with verbatim formatting.

Call centre. Observe.AI and CallMiner deliver full-conversation analytics, QA scoring and real-time agent guidance, increasingly fed by the conversational STT models above.

Accessibility. Ava offers AI captioning at ~95% accuracy or AI-plus-human at 99%, and Google Live Transcribe provides free on-device Android captioning in 70+ languages.

Feature comparison: the full matrix

Use-case specific recommendations

For most meetings

Winner: Fathom (free)

Unlimited recordings and transcripts at no cost, the highest G2 rating in the category, and fast summaries. Alternative: Granola if you want a bot-free, on-device notetaker on Mac.

For sales teams

Winner: Fireflies.ai ($10/user/month)

CRM auto-sync to Salesforce and HubSpot, 100+ languages and searchable meeting history. Alternative: tl;dv for deeper Zapier automation and global teams.

For privacy-sensitive conversations

Winner: Granola or local Whisper

Granola records on-device with no bot; for full data sovereignty, run NVIDIA Parakeet or faster-whisper locally so audio never leaves your machine. Alternative: Tactiq for a bot-free browser transcript that stores no audio.

For content creators

Winner: Descript ($16/month)

Text-based video editing remains a genuine workflow shift. Alternative: Riverside for recording-plus-transcription on remote interviews.

For developers building products

Winner: Deepgram Nova-3 / Flux

Best latency, entity accuracy and turn-taking for voice agents, from $0.26/hour. Alternative: AssemblyAI Universal-3 Pro if you need built-in sentiment, summaries and PII redaction.

For lowest cost at quality

Winner: Mistral Voxtral Mini Transcribe 2 ($0.003/minute)

Roughly 4% FLEURS WER at about a fifth of the price of premium APIs, with diarisation and 3-hour files. Alternative: self-hosted Parakeet or Groq-hosted Whisper at $0.04/hour for high-volume batch.

For highest accuracy

Winner: contested — test on your audio

Microsoft MAI-Transcribe-1 claims the lowest clean-English WER (3.8% FLEURS), with gpt-4o-transcribe, ElevenLabs Scribe v2 and Deepgram Nova-3 within two points. On clean audio the differences are marginal; on your real audio they are not. Alternative: Rev human transcription ($1.99/minute) when 99%+ is non-negotiable.

For enterprise and compliance

Winner: your incumbent cloud

Microsoft MAI-Transcribe-1 (Azure), Amazon Transcribe (AWS, HIPAA-eligible Medical) or Google Chirp 3 (GCP) — pick the one inside your existing compliance boundary. Alternative: Deepgram or AssemblyAI enterprise with on-prem options.

Pricing comparison: what you’ll actually pay

Consumer and creator tools (per month)

Developer APIs (per audio hour, list)

Cost strategy: route by job. Use a premium streaming model (Deepgram Flux, Scribe v2 Realtime) only for the live conversational layer, push post-call analytics to cheap batch (Voxtral, hosted Whisper, Whisper Turbo), and self-host Parakeet for high-volume, latency-tolerant work. The blended rate, not the headline number, is what moves your bill at scale.

What users actually think

Bot fatigue is real, and bot-free is the answer

Multiple notetaking bots joining the same call, awkward “who’s that recorder?” moments and consent worries have pushed users toward on-device tools. Granola’s surge to a $1.5 billion valuation and the growth of browser-extension tools like Tactiq track directly to this fatigue — the value proposition is “AI notes without a bot in the room.”

The Otter lawsuit hardened privacy sentiment

The class action crystallised long-simmering concerns; IT teams report network-wide blocks of auto-joining bots, and “get consent first” has become standard advice. For sensitive business conversations, users increasingly treat visible-bot, cloud-stored transcription as a compliance liability rather than a convenience.

Local Whisper earns consistent praise

Technical users favour local models for privacy and cost. faster-whisper and MacWhisper draw repeated praise for speed on Apple Silicon and “set it and forget it” privacy, with the main caveat being occasional hallucinations on noisy or silent stretches — now partly addressed by newer open models like Parakeet and Voxtral.

Descript: love the paradigm, watch the price

Text-based editing is still described as transformative for podcasters, but the 2025 pricing restructure drew sharp criticism from long-time users who felt existing value was cut. The workflow keeps them; the billing changes are the recurring complaint.

Recent launches reshaping the market (2026)

Microsoft MAI-Transcribe-1 (April 2026). Microsoft’s first proprietary speech model — not built on OpenAI infrastructure — claiming 3.8% FLEURS WER across 25 languages at roughly half the GPU cost, inside Azure AI Foundry (Microsoft).

OpenAI GPT-Realtime-Whisper (May 2026). A dedicated low-latency streaming transcription model at $0.017/minute, the first time OpenAI split streaming-optimised STT from the batch Whisper line.

Deepgram Flux Multilingual (April 2026). The first conversational speech-to-text with integrated end-of-turn detection, removing the external VAD layer and cutting 200–600ms off voice-agent response time.

Mistral Voxtral Transcribe 2 (February 2026). Two next-generation models — a $0.003/minute batch model with category-leading price-performance and an open-weight (Apache 2.0) sub-200ms realtime model for on-device use (Mistral).

AssemblyAI Universal-3 Pro (February 2026). A “speech language model” with natural-language keyterm prompting up to 1,500 words, bringing prompt-style control to transcription.

Granola’s $1.5B round (March 2026). The bot-free notetaker’s Series C confirmed on-device, no-bot capture as the consumer category’s centre of gravity (TechCrunch).

The Otter.ai class action consolidated. The case is now a single federal proceeding in San Jose, with the wider AI-notetaker industry’s recording practices effectively on trial (UC Today).

Frequently asked questions

Which AI transcription tool is most accurate in 2026?

On clean English audio the top providers cluster within about two percentage points of each other, so there is no single “most accurate” tool. Microsoft MAI-Transcribe-1 claims the lowest FLEURS word error rate at 3.8% across 25 languages; Deepgram Nova-3 (5.26% batch), AssemblyAI Universal-3 Pro (5.6%), OpenAI gpt-4o-transcribe (4.1%) and ElevenLabs Scribe v2 are all close behind. These are vendor numbers on clean benchmarks — the accuracy that predicts your results is WER measured on your own audio, where accents, code-switching and proper nouns separate the field.

What is the best free AI transcription tool?

Fathom is the best free meeting transcription tool, with unlimited recordings and transcripts and the highest G2 rating in the category. For privacy, faster-whisper and other local Whisper builds are free and keep audio on your machine. tl;dv also offers unlimited free recording, though free transcripts auto-delete after three months.

Is Otter.ai safe to use?

Use it only with explicit consent from everyone on the call. Otter.ai is the defendant in a consolidated federal class action alleging its OtterPilot bot recorded Zoom, Teams and Meet meetings without participants’ consent, with analysts estimating a possible $75–200 million settlement. The tool itself works well, but its auto-join behaviour has prompted IT blocks. For sensitive conversations, prefer bot-free tools like Granola or Tactiq, or local Whisper.

What is the cheapest speech-to-text API?

Mistral Voxtral Mini Transcribe 2 is the cheapest commercial API at $0.003/minute ($0.18/hour) with roughly 4% FLEURS word error rate — the best price-performance available. For lower still, self-hosted NVIDIA Parakeet costs only GPU time, and hosted Whisper on Groq runs about $0.04/hour. Among premium APIs, AssemblyAI ($0.21/hour async) and Deepgram Nova-3 ($0.26/hour batch) are the best value.

Which speech-to-text API is best for voice agents?

Deepgram Nova-3 and its Flux Multilingual model lead for voice agents, because Flux adds integrated end-of-turn detection that removes the external voice-activity-detection layer and saves 200–600ms of response time. ElevenLabs Scribe v2 Realtime and OpenAI GPT-Realtime-Whisper both deliver sub-150ms latency as alternatives. For voice agents, end-of-turn detection latency usually matters more than raw word error rate.

Can I run AI transcription locally for privacy?

Yes, and the open models are now genuinely competitive. NVIDIA Parakeet-TDT-0.6B-v3 tops the open-source leaderboard at 6.34% average WER across 25 European languages, Mistral’s Voxtral Realtime ships open weights under Apache 2.0 for sub-200ms on-device use, and faster-whisper, whisper.cpp and MacWhisper let you run Whisper on your own hardware. Local transcription keeps audio entirely on your machine, which removes the consent and data-residency risks of cloud tools.

How much does AI transcription cost for a business?

Meeting tools run $0–20/user/month — Fathom is free with unlimited recordings, while Fireflies ($10/user/month) and Otter ($8.33/month) charge per seat. Developer APIs run roughly $0.18–1.44/hour: for 100 hours a month, expect about $18 on Mistral Voxtral, $26 on Deepgram, or $144 on Amazon Transcribe. Self-hosting open models costs only hardware and GPU time, which wins at high volume.

What is word error rate (WER)?

Word error rate is the share of words a transcription system gets wrong — through substitutions, insertions or deletions — so lower is better. A 5% WER means roughly 1 word in 20 is incorrect, close to a careful human transcriber on clean audio. WER rises sharply on accents, background noise, technical vocabulary and code-switching, which is why a model’s clean-benchmark WER often overstates how it will perform on your real recordings.

Conclusion: how to choose in June 2026

AI transcription has matured past the accuracy race. On clean English the leading models are effectively tied, so the decision now turns on workflow, latency, privacy and cost.

• Best free meeting tool: Fathom — unlimited recordings and transcripts.
• Best bot-free notetaker: Granola — on-device, no bot in the call.
• Best for sales: Fireflies.ai ($10/user/month) — CRM sync and search.
• Best developer API: Deepgram Nova-3 / Flux — voice-agent latency and entity accuracy.
• Best value API: Mistral Voxtral Mini Transcribe 2 ($0.003/minute) — about a fifth of premium pricing.
• Best open-source / local: NVIDIA Parakeet-TDT-0.6B-v3 and Voxtral Realtime — competitive accuracy, full data sovereignty.
• Best for human accuracy: Rev ($1.99/minute) — 99%+ guaranteed.

The two non-negotiables: get consent before recording — the Otter litigation makes that a legal point, not just an etiquette one — and test two or three options on your own audio before you commit, because the only word error rate that matters is the one you get on your recordings, not the one in the launch post.

For related guides, see best AI for research, best AI voice generators, best AI voice cloning and the best AI apps.

This guide is updated as new transcription models launch and benchmarks evolve. Word error rates vary by audio profile and harness — vendor-reported figures run above real-world results, and we cite the source for each. Pricing and availability are subject to change.