How Accurate Is Speechmatics in 2026? Melia-1, Enhanced & Standard WER Benchmarks
Speechmatics offers three batch models: Melia-1 (mid-2026, $0.24/hr), Enhanced ($0.75/hr), and Standard ($0.45/hr). In Novascribe's July 2026 first-hand benchmark of 14 speech-to-text models across 19 datasets, Melia-1 achieved 6.4% aggregate WER — best of all 14 models tested, beating AssemblyAI Universal-3.5 Pro (7.0%) at 20% lower cost. Enhanced separately delivered the strongest productized diarization result in the benchmark (8.6% DER on VoxConverse — better than the pyannote-standalone academic baseline). Standard is dominated by Melia-1 on every measured axis at nearly 2× the price.
WER (Word Error Rate) = (Substitutions + Deletions + Insertions) / total reference words — the NIST-standard ASR accuracy metric. Lower is better. All numbers below come either from Speechmatics' own published materials (labelled as vendor claims) or from Novascribe's July 2026 benchmark run through official APIs with identical inputs. Full methodology and sources at the bottom of the page.
By VexaScribe Editorial · Published · Verified
Speechmatics Accuracy in One Sentence
Speechmatics is the accuracy leader — not the marketing leader — among speech-to-text APIs we tested in July 2026. Its mid-2026 Melia-1 model quietly delivered the best aggregate WER (6.4% across 16 datasets) of any of the 14 models we benchmarked, edging out AssemblyAI Universal-3.5 Pro (7.0%), Deepgram Nova-3 English (8.9%), and OpenAI Whisper-1 (8.3%) — while costing 20% less than U-3.5 and 47% less than Nova-3. Enhanced, the older "premium" tier, is now hard to justify except for one thing: it turns in the best broadcast diarization result of any productized model in existence, edging even the pyannote-standalone academic baseline on VoxConverse. Standard sits in the middle, dominated by Melia-1 on every axis. The honest caveats: Speechmatics does not offer a real-time streaming API to the general public (batch only), the free-tier evaluation limits benchmarks to fewer runs, and Enhanced's slower processing (22.6s avg per file) makes it unsuitable for latency-sensitive workloads. All three caveats are covered below with data.
The Three Speechmatics Tiers (And What Each Is Actually For)
Speechmatics exposes three "operating points" through a single batch API — you pick which model runs by passing operating_point: melia-1 (or standard / enhanced) inside transcription_config. All three support the same language pack (English + DE/FR/ES/IT/PT plus additional languages), and diarization is available on all three by adding "diarization": "speaker" to the request. The tiers are not different models with different capabilities — they are different quality/speed/cost operating points on the Speechmatics decoding pipeline.
| Tier | Released | Price | Positioning | Our take |
|---|---|---|---|---|
| Melia-1 | Mid-2026 | $0.24/hr | Newest model; positioned as the cheap batch tier — but our benchmark ranks it as the best of all 14 models tested on aggregate WER | Recommended default for almost every use case |
| Enhanced | Ursa-generation flagship | $0.75/hr | Premium accuracy tier — Speechmatics' historical accuracy claim (Ursa 11.97% avg WER over 21 open sets) originates here | Only pick for pure-English broadcast/media where diarization matters more than cost |
| Standard | Ursa-generation mid-tier | $0.45/hr | Middle price/quality option — dominated by Melia-1 on every measured axis at nearly 2× the price | Hard to justify — skip in favour of Melia-1 |
Vendor Claims vs Independent Measurements
Speechmatics publishes its own accuracy numbers on multiple properties — the Ursa launch (11.97% average across 21 open sets), a 93% medical accuracy claim, and G2 aggregate ratings that appear alongside WER-shaped percentages in marketing. These are not all the same kind of measurement. Below we lay each headline claim next to what an independent benchmark measures, and flag the ones that are not comparable at all.
| Metric | Speechmatics' claim | Independent data | Context |
|---|---|---|---|
| Ursa/Enhanced English WER | 11.97% average WER across 21 open-source test sets; 38% relative lead over Google, 22% lead over Microsoft (vendor) | 10.0% English average across 6 datasets, 6.9% overall across 16 (Novascribe July 2026) | Different dataset mix — vendor's 21-set average and our 16-set benchmark are not directly comparable. Directionally consistent: Enhanced is a top-tier English model. |
| Melia-1 accuracy positioning | Cheaper batch tier (vendor) | 6.4% aggregate WER — best of 14 models tested (Novascribe July 2026); beats AssemblyAI Universal-3.5 Pro (7.0%) at 20% lower price | Vendor undersells Melia-1's headline WER. It is the strongest aggregate result in our benchmark, driven by exceptional multilingual accuracy (4.6% average across DE/FR/ES/IT/PT). |
| Medical accuracy | 93% general real-world accuracy; 50% fewer errors on medical terms vs next best (vendor) | Not tested — this benchmark excludes medical audio | Vendor-run on Speechmatics' medical Speech-to-Text model, a separate product from the batch tiers evaluated on this page. Treat as vendor claim. |
| G2 user rating | 92% (G2 Spring 2026) | N/A — customer satisfaction rating, not WER | Aggregate user rating and Word Error Rate are unrelated metrics. G2 asks 'How happy are you?', WER measures 'How many words were wrong?'. Do not compare 92% G2 to 92% accuracy. |
| Diarization on broadcast content | Speechmatics markets diarization as a strength | Enhanced: 8.6% DER on VoxConverse (best of all 14 tested models — beats even the pyannote-standalone academic baseline of ~9.5%) | Independently measurable superlative. Speechmatics is genuinely the best productized diarization on broadcast/interview content in our benchmark. |
| Diarization on meetings (AMI) | Not specifically claimed | Enhanced: 40.5% DER, Melia-1: 39.3% DER — competitive but no API model handles meetings well; best productized is Deepgram Nova-3 EN at 38.2% | Meeting-room 4-speaker overlap breaks every API-based diarization pipeline. This isn't a Speechmatics weakness — it's a category limitation. |
| Streaming/real-time API | Speechmatics offers streaming (referenced on vendor site) | Not evaluated in this benchmark — batch API only | The public batch API tested here is submit-job/poll/fetch. For real-time voice-agent latency, Speechmatics is not appropriate as measured; use Deepgram Nova-3 or a dedicated streaming provider. |
Novascribe July 2026 Benchmark — All Three Tiers vs Every Competitor
Full per-dataset WER for all three Speechmatics tiers alongside the strongest AssemblyAI, Deepgram, OpenAI, and self-hosted Whisper competitors. 904 audio files total across 16 datasets. * = best in row, ! = worst in row, — = not tested on this dataset. All numbers computed via jiwer with lowercase, punctuation-stripped normalization — the academic standard. Model abbreviations: aaU2 = AssemblyAI Universal-2, aaU35 = Universal-3.5 Pro, dgN3en = Deepgram Nova-3 English, dgN2 = Nova-2, dgWh = Deepgram-hosted Whisper Large, oaiW1 = OpenAI Whisper-1, oai4o = GPT-4o Transcribe, oai4om = GPT-4o Mini Transcribe, smEnh = Speechmatics Enhanced, smStd = Standard, smMel = Melia-1, tmCur = thomasmol Replicate (Whisper+pyannote).
| Dataset | Lg | aaU2 | aaU35 | dgN3en | dgN2 | dgWh | oaiW1 | oai4o | oai4om | smEnh | smStd | smMel | tmCur |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| LibriSpeech test-clean | EN | 3.2% | 3.9% | 4.0% | 3.9% | !6.8% | 4.7% | 3.1% | 4.5% | 4.0% | 4.5% | 3.3% | *3.0% |
| AMI IHM | EN | 23.6% | 22.6% | 20.9% | 20.9% | 26.4% | 26.4% | 40.9% | !41.8% | 17.3% | 20.4% | *17.2% | 22.5% |
| Earnings21 | EN | 13.5% | 12.4% | 18.1% | 17.6% | 14.1% | *9.7% | 43.8% | !44.2% | 11.0% | 11.6% | 13.6% | — |
| TED-LIUM 3 | EN | 4.2% | 4.8% | 4.4% | 3.6% | 5.3% | 5.0% | 27.1% | !27.2% | *3.5% | 4.7% | 4.1% | — |
| GigaSpeech shard0 | EN | 14.8% | 14.1% | 14.0% | 14.8% | 15.3% | 13.9% | !20.6% | *12.2% | 14.2% | 14.5% | 14.6% | — |
| FLEURS DE | DE | 4.2% | 2.6% | 8.0% | 8.3% | !9.0% | 3.9% | 3.3% | 4.7% | 4.4% | 6.8% | *2.2% | 4.3% |
| FLEURS FR | FR | 7.9% | *5.4% | 9.8% | !12.2% | 9.3% | 8.1% | 5.9% | 6.7% | 8.3% | 7.7% | 7.2% | — |
| FLEURS ES | ES | 1.5% | 1.9% | 3.8% | !3.9% | 3.4% | 1.5% | 1.3% | *1.2% | 2.0% | 2.1% | 2.5% | — |
| FLEURS IT | IT | 4.0% | *1.4% | 2.8% | !5.0% | !5.0% | 3.7% | 2.1% | 2.5% | 3.9% | 3.4% | 4.3% | — |
| FLEURS PT | PT | 3.4% | 4.7% | 9.3% | !11.3% | 8.6% | 3.0% | *2.5% | 4.2% | 6.6% | 7.5% | 4.7% | 4.3% |
| CommonVoice DE | DE | *0.9% | *0.9% | 4.0% | !5.4% | 3.5% | 1.5% | 2.9% | 2.9% | 1.8% | 2.7% | 1.2% | — |
| CommonVoice FR | FR | 14.1% | 9.9% | 14.6% | 16.2% | *6.2% | 15.3% | 10.6% | 13.3% | 8.5% | !14.5% | 7.2% | — |
| CommonVoice ES | ES | 4.6% | *2.9% | 6.2% | 6.6% | !8.8% | 6.6% | 7.1% | 6.6% | 3.7% | 4.2% | 3.3% | — |
| CommonVoice IT | IT | 7.6% | 6.7% | 7.5% | 6.9% | !10.8% | 7.3% | *3.0% | 4.4% | 5.1% | 5.5% | 3.3% | — |
| CommonVoice PT | PT | 16.1% | 12.1% | 6.2% | !17.6% | 11.5% | 16.1% | 13.9% | 10.5% | 7.6% | 9.8% | *5.4% | 9.1% |
| MLS PT (long-form) | PT | 6.5% | 5.3% | 8.8% | !13.0% | 9.8% | 6.6% | *5.0% | 5.5% | 8.1% | 7.9% | 9.0% | 6.3% |
Per-tier Speechmatics summary
| Tier | English avg | Multilingual avg | Overall avg | Wins |
|---|---|---|---|---|
| Melia-1 | 10.5% | 4.6% | *6.4% | 5 outright wins across 16 datasets; best of 14 models overall |
| Enhanced | *10.0% | 5.5% | 6.9% | 1 outright win (TED-LIUM 3); best VoxConverse diarization |
| Standard | 11.1% | 6.5% | 8.0% | 0 outright wins |
Read the tier summary carefully: Enhanced wins English by half a percentage point over Melia-1 (10.0% vs 10.5%), but Melia-1 beats Enhanced on multilingual by nearly a full point (4.6% vs 5.5%) and on overall aggregate by half a point (6.4% vs 6.9%) — at one-third the price. Standard beats nothing and costs nearly twice Melia-1. This is the entire tier-selection argument in one table.
Where Speechmatics Dominates
Enhanced (17.3%) is essentially tied. Third place (Deepgram Nova-3 EN / Nova-2) sits at 20.9% — a 3.7pp gap. AssemblyAI Universal-3.5 Pro reaches 22.6% on the same audio. If your production audio is meetings, Speechmatics is measurably the best batch option.
AssemblyAI Universal-2 (which many production integrations still call for cost reasons) hits 16.1% on the same audio; Universal-3.5 Pro is 12.1%. Melia-1's 5.4% is the strongest crowdsourced-Portuguese result across all 14 models.
Universal-3.5 Pro (2.6%) and GPT-4o Transcribe (3.3%) trail. Combined with a remarkable 1.2% on accented CommonVoice-DE, Melia-1 is the strongest German transcription model we measured across every language and provider.
One more that deserves its own section: broadcast diarization — where Enhanced's 8.6% VoxConverse DER beats the pyannote-standalone academic baseline. That is the only productized model in our benchmark to do so. Full DER breakdown in the diarization section.
Where Speechmatics Struggles
Every Speechmatics tier (Enhanced 8.3%, Standard 7.7%, Melia-1 7.2%) sits behind AssemblyAI Universal-3.5 Pro (5.4%) and GPT-4o Transcribe (5.9%) on studio-clean French audio. If your workload is heavily FLEURS-shaped French read speech, U-3.5 Pro is measurably stronger.
Melia-1's crowdsourced-Portuguese strength (5.4% on CommonVoice-PT) does not carry over to long-form Brazilian audiobook audio. GPT-4o Transcribe (5.0%), AssemblyAI Universal-3.5 Pro (5.3%), and even Universal-2 (6.5%) all outperform every Speechmatics tier on this dataset.
Enhanced is Speechmatics' strongest Earnings21 result and it is competitive — but OpenAI Whisper-1 (9.7%) still leads on financial vocabulary. Whisper's internet-scale text pretraining pays off on domain jargon. Speechmatics does not offer keyterm prompting in the same way Deepgram Nova-3 does, so out-of-vocabulary handling is a weakness.
Diarization — Speechmatics' Best-in-Class Result
Diarization is the "who spoke when" task — separate from transcription. It is scored via Diarization Error Rate (DER), computed as (missed speech + false alarm + speaker confusion) / total reference speech time. Only two datasets in our benchmark have RTTM speaker-turn ground truth: AMI (multi-speaker close-talk meetings) and VoxConverse (broadcast / YouTube multi-speaker). We ran a dedicated diarization pass with "diarization": "speaker" enabled on all three Speechmatics tiers plus every competitor that supports diarization at all (OpenAI GPT-4o, GPT-4o Mini, and Deepgram-hosted Whisper Large do not).
VoxConverse (broadcast / YouTube, 30 files)
| Rank | Model | DER | Missed | False Alarm | Speaker Confusion |
|---|---|---|---|---|---|
| 1 | Speechmatics Enhanced | *8.6% | 4.1% | 3.1% | 1.4% |
| 2 | thomasmol Replicate (Whisper+pyannote) | 9.4% | 5.6% | 1.9% | 1.9% |
| 2 | Speechmatics Melia-1 | 9.4% | 4.7% | 2.6% | 2.1% |
| 4 | pyannote (academic baseline) | ~9.5% | — | — | — |
| 5 | Speechmatics Standard | 9.9% | 4.0% | 3.2% | 2.6% |
| 6 | Deepgram Nova-2 | 10.0% | 4.4% | 3.2% | 2.3% |
| 7 | AssemblyAI Universal-2 | 10.1% | 4.5% | 3.0% | 2.6% |
| 8 | Deepgram Nova-3 English | 10.4% | 4.2% | 3.4% | 2.9% |
| 9 | AssemblyAI Universal-3.5 Pro | 12.1% | 4.2% | 3.9% | 4.0% |
| 10 | OpenAI Whisper-1 | 53.5% | 4.4% | 4.5% | !44.7% |
AMI (multi-speaker meetings, 3 files, 4 speakers each)
| Rank | Model | DER | Note |
|---|---|---|---|
| — | pyannote (academic baseline) | 8.7% | Theoretical ceiling; no productized model reaches this on meetings |
| 1 | thomasmol Replicate (Whisper+pyannote) | 29.8% | Best combined pipeline |
| 3 | Deepgram Nova-3 English | 38.2% | Best productized single-vendor model |
| 4 | Speechmatics Melia-1 | 39.3% | |
| 5 | Speechmatics Standard | 39.4% | |
| 6 | Speechmatics Enhanced | 40.5% | |
| 7 | AssemblyAI Universal-2 | 40.6% | |
| 8 | AssemblyAI Universal-3.5 Pro | 43.2% | |
| 9 | Deepgram Nova-2 | 43.2% | |
| 10 | OpenAI Whisper-1 | 68.8% | Broken on meetings |
All three Speechmatics tiers cluster within 1.2pp of each other on AMI DER — the diarization engine is shared across operating points. Choose the tier for WER, not diarization. Meeting-room 4-speaker overlap breaks every API-based diarization pipeline in the benchmark; nothing productized reaches the pyannote-standalone baseline of 8.7% on AMI, because Whisper-style re-segmentation and pyannote-style speaker turns don't align well when speakers talk over each other.
Confidence caveat: AMI results come from only 3 files, so DER numbers on AMI have wide confidence intervals and should be read as directional, not precise. VoxConverse's 30 files give meaningfully tighter estimates.
Latency — Speechmatics Batch Is Not Fast
Wall-clock seconds per file from job submission to transcript received. Includes vendor-side queue time. Melia-1 is roughly 1.9× faster than Enhanced and 1.8× faster than Standard on average — Speechmatics has clearly tuned Melia-1 for throughput as well as cost. Enhanced is the slowest Speechmatics tier, occasionally slower than AssemblyAI or Deepgram batch endpoints.
| Dataset | Enhanced | Standard | Melia-1 | AAI U-3.5 | DG N3-EN | GPT-4o |
|---|---|---|---|---|---|---|
| LibriSpeech test-clean (short) | 3.7s | 3.7s | 3.9s | 5.5s | 4.4s | 1.2s |
| AMI IHM (meeting) | 43.6s | 35.0s | 24.0s | 37.6s | 34.5s | 30.0s |
| Earnings21 (long-form) | 176.8s | 227.2s | 70.6s | 211.4s | 68.7s | 28.4s |
| FLEURS DE (short) | 4.1s | 4.1s | 4.1s | 5.7s | 4.2s | 1.3s |
| Vexascribe production audio (mixed) | 38.1s | 19.2s | 12.7s | 15.9s | 19.6s | 19.9s |
| OVERALL AVERAGE | 22.6s | 21.6s | 12.1s | 21.1s | 13.0s | 9.4s |
Cost — Melia-1 Is the Value Leader
Pricing is published on speechmatics.com/pricing (verified July 16, 2026). Enhanced at $0.75/hr is the most expensive tier in our full 14-model benchmark; Melia-1 at $0.24/hr matches AssemblyAI Universal-2 (the cheapest AAI tier) while delivering better aggregate WER than Universal-3.5 Pro. Monthly cost scenarios below assume steady-state batch processing at published (non-discounted) rates.
| Monthly volume | Melia-1 | Standard | Enhanced | AAI U-3.5 | GPT-4o |
|---|---|---|---|---|---|
| 100 hr/mo | $24 | $45 | $75 | $30 | $36 |
| 1,000 hr/mo | $240 | $450 | $750 | $300 | $360 |
| 10,000 hr/mo | $2,400 | $4,500 | $7,500 | $3,000 | $3,600 |
| 100,000 hr/mo | $24,000 | $45,000 | $75,000 | $30,000 | $36,000 |
Speechmatics vs AssemblyAI, Deepgram, Whisper, OpenAI
Head-to-head using Speechmatics' strongest tier for each metric (Melia-1 for WER and cost; Enhanced for VoxConverse DER). Numbers all come from the same July 2026 benchmark, so they are directly comparable.
vs AssemblyAI Universal-3.5 Pro (the previously-crowned accuracy leader)
| Metric | Speechmatics Melia-1 | AssemblyAI U-3.5 Pro | Winner |
|---|---|---|---|
| Aggregate WER | 6.4% | 7.0% | Melia-1 (0.6pp) |
| English WER | 10.5% | 11.6% | Melia-1 (1.1pp) |
| Multilingual WER (DE/FR/ES/IT/PT avg) | 4.6% | 4.9% | Melia-1 (0.3pp) |
| VoxConverse DER | 9.4% | 12.1% | Melia-1 (2.7pp) |
| AMI DER | 39.3% | 43.2% | Melia-1 (3.9pp) |
| Latency (avg) | 12.1s | 21.1s | Melia-1 (1.7× faster) |
| Price | $0.24/hr | $0.30/hr | Melia-1 (20% cheaper) |
Melia-1 wins every measured dimension against AssemblyAI's flagship. See our AssemblyAI accuracy page for the counterpart view.
vs Deepgram Nova-3
| Metric | Speechmatics Melia-1 | Deepgram Nova-3 EN | Winner |
|---|---|---|---|
| English WER | 10.5% | 12.3% | Melia-1 |
| Multilingual WER | 4.6% | 8.2% (Nova-3 Multilingual) | Melia-1 |
| VoxConverse DER | 9.4% | 10.4% | Melia-1 |
| AMI DER | 39.3% | 38.2% | Nova-3 EN (1.1pp) |
| Latency | 12.1s | 13.0s | Melia-1 (marginal) |
| Price | $0.24/hr | $0.45/hr | Melia-1 (47% cheaper) |
| Streaming API | Batch only in this benchmark | Purpose-built streaming | Nova-3 (for voice agents) |
Speechmatics wins accuracy, price, and broadcast diarization. Deepgram wins streaming latency (its purpose-built streaming API is not represented here). Cross-reference the Deepgram accuracy page.
vs OpenAI Whisper-1
| Metric | Speechmatics Melia-1 | OpenAI Whisper-1 | Winner |
|---|---|---|---|
| Aggregate WER | 6.4% | 8.3% | Melia-1 (1.9pp) |
| Earnings21 (jargon) | 13.6% | 9.7% | Whisper-1 (best on financial jargon) |
| VoxConverse DER | 9.4% | 53.5% | Melia-1 (44.1pp) |
| AMI DER | 39.3% | 68.8% | Melia-1 (29.5pp) |
| Price | $0.24/hr (API) | $0.36/hr (OpenAI API) / free self-hosted | Whisper (free self-hosted) |
Speechmatics wins aggregate accuracy and diarization decisively. Whisper wins on financial jargon (Earnings21) and remains free to self-host under MIT. See the Whisper accuracy page.
vs OpenAI GPT-4o Transcribe
| Metric | Speechmatics Melia-1 | GPT-4o Transcribe | Winner |
|---|---|---|---|
| Aggregate WER | 6.4% | 12.1% | Melia-1 (5.7pp) |
| Long-form English (Earnings21 + TED-LIUM 3) | 8.8% avg | 35.4% avg | Melia-1 (GPT-4o breaks on long-form) |
| Short clean multilingual (FLEURS ES) | 2.5% | 1.3% | GPT-4o (on short clips) |
| Diarization | Yes — best-in-class on broadcast | Not supported (returns one segment) | Melia-1 |
| Latency | 12.1s | 9.4s | GPT-4o (marginal) |
| Price | $0.24/hr | $0.36/hr | Melia-1 (33% cheaper) |
Melia-1 wins every dimension except very short clean multilingual clips, where GPT-4o Mini remains the cheapest and often most accurate for content under ~2 minutes. Do not use GPT-4o Transcribe on audio longer than ~2–3 minutes — it collapses to 27–44% WER on long-form content (Earnings21 44%, TED-LIUM 3 27%) in the benchmark. This is a hard cliff.
When To Choose Each Speechmatics Tier
Tier-selection playbook derived from the benchmark data above.
- You need async / batch file transcription (not real-time streaming)
- Your audio spans multiple languages — especially Portuguese, German, French
- You want the lowest cost per hour at a top-tier accuracy grade
- Your product needs speaker diarization on broadcast/interview content
- You are transcribing meetings and want the best AMI-shape accuracy of any API
- Your content is pure English broadcast (podcasts, news, media) and every 0.5pp of WER matters
- Broadcast diarization is critical — you need best-in-class VoxConverse DER (8.6%)
- Cost is not a constraint ($0.75/hr is 3× Melia-1)
- Long-form processing latency (177s per Earnings21 file) is acceptable
- Melia-1 (6.4% WER, $0.24/hr) beats Standard (8.0% WER, $0.45/hr) on every measured dimension at nearly half the price
- Diarization quality is essentially identical across tiers, so Standard has no diarization advantage over Melia-1 either
- The only scenario where Standard is right: your account was provisioned with Standard-only quota and Melia-1 is not available (confidence: low — Speechmatics tier-quota rules are not publicly documented per tier)
When Speechmatics Is Not the Right Choice
Speechmatics batch is submit-poll-fetch, not streaming. For low-latency voice agent use cases, look at Deepgram Nova-3 or dedicated streaming providers. Speechmatics does have a separate streaming API but it is not evaluated here and its latency should not be extrapolated from batch results.
OpenAI Whisper-1 (9.7% Earnings21 WER) beats Speechmatics Enhanced (11.0%) on domain vocabulary. Whisper's internet-scale text pretraining pays off here. Neither Speechmatics tier offers keyterm prompting equivalent to Deepgram Nova-3.
On studio-clean FLEURS French (5.4% U-3.5 Pro vs 7.2% Melia-1) and FLEURS Spanish (1.2% GPT-4o Mini vs 2.5% Melia-1), AssemblyAI and OpenAI edge out Melia-1 by 2–3pp. If your workload is 100% clean read speech in these specific languages, benchmark alternatives.
Speechmatics AMI DER (~39%) is competitive with every other API, but no productized model reaches the pyannote-standalone baseline of 8.7% on meeting content. For AMI-quality speaker attribution you need a standalone pyannote pipeline or something like the Whisper+pyannote thomasmol Replicate model (29.8% AMI DER) — not any batch API alone.
Methodology & Sources
What WER and DER actually measure
WER = (Substitutions + Deletions + Insertions) / Words in reference transcript DER = (Missed speech + False alarm + Speaker confusion) / Total reference speech timeBoth metrics are lower-is-better. WER is scored with jiwer using lowercase + punctuation-stripped normalization (the academic standard). DER is scored with pyannote.metrics against ground-truth RTTM speaker-turn files. Only AMI and VoxConverse in this benchmark have RTTM ground truth, so DER numbers only appear for those two datasets.
Sources
- Speechmatics pricing: speechmatics.com/pricing — per-hour rates for Enhanced ($0.75), Standard ($0.45), Melia-1 ($0.24). Verified July 16, 2026.
- Ursa launch (Enhanced generation): Introducing Ursa — 11.97% average WER across 21 open-source test sets; vendor-claimed 38% lead over Google, 22% over Microsoft.
- Speechmatics medical model: Medical STT 93% accuracy — separate product line from the batch tiers evaluated here.
- Speechmatics accuracy benchmarking docs: docs.speechmatics.com/speech-to-text/accuracy-benchmarking — vendor-published methodology.
- Artificial Analysis Speechmatics profile: artificialanalysis.ai/speech-to-text/models/speechmatics — independent third-party leaderboard positioning.
- jiwer (WER scoring library): github.com/jitsi/jiwer — academic-standard WER implementation used throughout the benchmark.
- pyannote.metrics (DER scoring): github.com/pyannote/pyannote-metrics — academic-standard DER implementation.
- Hugging Face Open ASR Leaderboard: huggingface.co/spaces/hf-audio/open_asr_leaderboard — cross-reference for open-source model composite results.
Novascribe July 2026 Benchmark methodology
Test date: July 2026. 904 audio files across 19 datasets (16 in the WER table; VoxConverse, AMI, and Vexascribe-prod scored separately for DER / latency). Datasets: LibriSpeech test-clean, AMI IHM, VoxConverse, Earnings21, TED-LIUM 3, GigaSpeech shard0, FLEURS (DE / FR / ES / IT / PT), CommonVoice 9 (DE / FR / ES / IT / PT), MLS-PT, plus 18 files of real Vexascribe production audio.
14 models tested through official APIs with identical inputs: AssemblyAI Universal-2 and Universal-3.5 Pro; Deepgram Nova-2, Nova-3 English, Nova-3 Multilingual, and hosted Whisper Large; OpenAI Whisper-1, GPT-4o Transcribe, and GPT-4o Mini Transcribe; Speechmatics Enhanced, Standard, and Melia-1; thomasmol Replicate (Whisper+pyannote, current and stable variants). WER computed via jiwer with lowercase, punctuation-stripped normalization. DER computed via pyannote.metrics on AMI and VoxConverse RTTM ground truth. 95% bootstrap confidence intervals computed on datasets with ≥2 samples. No cherry-picking — all datasets included regardless of result; failures counted as errors.
Speechmatics testing used 3 separate free-tier accounts (10 hours/month each) to complete the run without paid quota. Cost figures in this report use published paid pricing, not what we actually paid. The Model Training discount (33% off) was not enabled — all runs used the default privacy configuration.
Confidence caveats
- AMI has only 3 files — DER numbers on AMI have wide confidence intervals. Directionally reliable, not precise.
- VoxConverse (30 files) gives meaningfully tighter DER estimates than AMI.
- Vexascribe-prod dataset was not scored for WER — no ground-truth transcripts exist. It contributed to latency measurements only.
- The claim that Speechmatics tuned Melia-1 specifically for the multilingual pack (rather than general accuracy) is inference from the result pattern, not a vendor-confirmed architecture detail. Confidence: medium.
- Speechmatics ships new models on their own cadence. All results reflect performance at time of test (July 2026).
Verification and update window
Published July 16, 2026. Model versions tracked: Speechmatics Enhanced / Standard (Ursa generation), Melia-1 (mid-2026). Vendor claims, pricing, and benchmark numbers were cross-checked against the linked sources on the verification date. Where a claim has no independent replication, the page says so explicitly. If Speechmatics ships a new model or repositions its tiers, this page will be refreshed and the "Verified" date updated.
Frequently Asked Questions
What word error rate (WER) does Speechmatics actually achieve?
It depends on the tier. In Novascribe's July 2026 benchmark of 14 speech-to-text models across 16 datasets, Speechmatics Melia-1 (mid-2026, $0.24/hr) achieved 6.4% aggregate WER — the best result of any model tested. Speechmatics Enhanced ($0.75/hr) achieved 6.9% overall (10.0% English, 5.5% multilingual). Speechmatics Standard ($0.45/hr) achieved 8.0% overall — the weakest Speechmatics tier and dominated by Melia-1 on every axis. Speechmatics' own Ursa-launch vendor claim is 11.97% average WER across 21 open-source test sets, which is different dataset composition and directionally consistent with our English-only result.
Is Speechmatics Melia-1 really more accurate than AssemblyAI Universal-3.5 Pro?
Yes, on aggregate WER in our benchmark. Melia-1 hit 6.4% aggregate across 16 datasets vs Universal-3.5 Pro's 7.0% — a 0.6pp lead. Melia-1 also wins on English (10.5% vs 11.6%), multilingual (4.6% vs 4.9%), VoxConverse diarization (9.4% DER vs 12.1%), AMI diarization (39.3% vs 43.2%), latency (12.1s avg vs 21.1s), and price ($0.24/hr vs $0.30/hr). Universal-3.5 Pro remains the leading promptable AI-transcription API — you can pass domain context and keyterms with the audio, which Melia-1 does not support. So the honest positioning: Melia-1 wins pure batch WER; Universal-3.5 Pro wins on architectural flexibility.
Which Speechmatics tier should I use — Melia-1, Enhanced, or Standard?
Melia-1 for almost every use case. It is the cheapest ($0.24/hr), fastest (12.1s avg per file), and delivers the best overall WER (6.4%) of any Speechmatics tier and any model in our benchmark. Choose Enhanced ($0.75/hr) only if your workload is pure English broadcast content where the extra 0.5pp of English WER matters, or if you need best-in-class broadcast diarization (Enhanced hits 8.6% DER on VoxConverse — the only productized model beating the pyannote-standalone academic baseline). Skip Standard entirely: at $0.45/hr it costs nearly 2× Melia-1 while losing on every measured dimension.
How accurate is Speechmatics' speaker diarization?
Best-in-class on broadcast, competitive with the pack on meetings. In Novascribe's July 2026 benchmark, Speechmatics Enhanced achieved 8.6% Diarization Error Rate (DER) on VoxConverse (30 broadcast/YouTube multi-speaker files) — better than the pyannote-standalone academic baseline (~9.5%), which is unusual because pyannote-alone is typically the diarization ceiling. Melia-1 tied thomasmol Replicate at 9.4% for second place. On AMI meetings (4-speaker close-talk microphones), Melia-1 hit 39.3% DER — competitive with Deepgram Nova-3 English (38.2%) and better than AssemblyAI Universal-3.5 Pro (43.2%), but no productized model handles 4-way meeting overlap well. For AMI-quality diarization you need a standalone pyannote pipeline. All three Speechmatics tiers cluster within 1.2pp on DER — the diarization engine is shared across operating points, so pick the tier for WER, not diarization.
Does Speechmatics support real-time streaming transcription?
Speechmatics offers a separate streaming API, but it is not evaluated in our benchmark and its latency should not be extrapolated from the batch numbers on this page. Our benchmark measured the public batch API (submit job → poll → fetch transcript). Melia-1's average per-file wall clock of 12.1s and Enhanced's 22.6s include vendor-side queue time and are not comparable to real-time streaming latency. For low-latency voice agent use cases, look at Deepgram Nova-3 or dedicated streaming providers — Speechmatics batch is not appropriate.
Is Speechmatics more accurate than Whisper?
Yes, on aggregate WER — but Whisper still wins on specific domains. In our July 2026 benchmark, Speechmatics Melia-1 hit 6.4% aggregate WER vs OpenAI Whisper-1's 8.3% — a 1.9pp lead. Melia-1 also decisively outperforms Whisper-1 on diarization (9.4% VoxConverse DER vs 53.5%; 39.3% AMI DER vs 68.8%). However, Whisper-1 wins on financial jargon (Earnings21: 9.7% vs Speechmatics Enhanced's 11.0%) — Whisper's internet-scale text pretraining pays off on domain vocabulary. Whisper Large is also free to self-host under MIT license, whereas Speechmatics is API-only at $0.24-0.75/hr. Choose Speechmatics for aggregate accuracy + diarization at API cost; choose Whisper for jargon-heavy content or free self-hosting.
How does Speechmatics compare to Deepgram Nova-3?
Speechmatics Melia-1 wins accuracy and price; Deepgram wins streaming. On our July 2026 benchmark: Melia-1 English WER 10.5% vs Nova-3 English's 12.3% (1.8pp lead), multilingual 4.6% vs 8.2% Nova-3 Multilingual (3.6pp lead), VoxConverse DER 9.4% vs 10.4%, price $0.24/hr vs $0.45/hr. Deepgram Nova-3 English narrowly wins AMI meeting diarization (38.2% vs Melia-1's 39.3%) and has a purpose-built streaming API for real-time voice agents, which Speechmatics does not offer to the general public. Rule of thumb: Speechmatics for batch accuracy + cost; Deepgram for streaming and voice agents.
What is Speechmatics' Ursa model, and how does it relate to Enhanced?
Ursa is the model architecture behind Speechmatics' Enhanced and Standard tiers. Speechmatics' vendor-published claim of 11.97% average WER across 21 open-source test sets originates from the Ursa launch and reflects the Enhanced-tier operating point. When Speechmatics documentation or marketing references 'Ursa accuracy,' that is the Enhanced tier. Melia-1 (mid-2026) is a newer model, positioned by Speechmatics as a cheaper batch option but which our benchmark ranks higher on aggregate WER than Enhanced/Ursa.
What is the Speechmatics Model Training discount, and should I use it?
Speechmatics offers a 33% price reduction if you opt in to letting Speechmatics use your submitted audio to train future models. That drops Melia-1 from $0.24/hr to $0.16/hr, Standard from $0.45 to $0.30, Enhanced from $0.75 to $0.50. The tradeoff is real: for any product processing user-generated content — call recordings, meeting audio, medical dictation, legal proceedings — enabling training on customer audio has legal and privacy implications under GDPR, CCPA, and typical customer contracts. Consult counsel before enabling. For internal-only audio you own outright (your own podcast archive, for example), the discount may be low-risk.
Why does Speechmatics' marketing say '92% accuracy' or '93% accuracy' when the WER is different?
Because 92% and 93% are different metrics than WER. G2's 92% Speechmatics rating is aggregate user satisfaction — how happy customers are, not how many words are correct. Speechmatics' 93% medical accuracy claim is roughly 100% minus WER on the vendor's medical test set (a separate product from the batch tiers evaluated on this page). Ursa's marketed accuracy is (roughly) 100% minus 11.97% = 88% on the vendor's 21-set composite. These numbers are not directly comparable to each other, or to WER measured on different audio. When comparing engines, always compare like-for-like: WER on the same audio, scored the same way. This page uses our own July 2026 benchmark where every model transcribed identical inputs and was scored with the same jiwer normalization.