Skip to content
audo

AI Speech Enhancement Benchmark 2026: 124 Clips Scored on Word Errors, Sound Quality & Speed

Reproducible measurements across the 100-clip Dolby UNIVERSE validation set (16 kHz) and 24 full-band 48 kHz recordings—testing non-generative neural masks against generative voice restoration on word preservation, ITU-T P.835 DNSMOS, speaker similarity, PESQ, ESTOI, bandwidth, throughput, and MP4 video passthrough.

Updated November 18, 2026Michael NguyenTested November 2026

Try Audio enhancer from Audo

Full

From subtle to full.

More options
Free for the first minute, 3 times a day. No sign-up.

Quick answer

Across 100 Dolby UNIVERSE clips (16 kHz) and 24 full-band 48 kHz recordings, no single model wins both background noise and room echo—which is why Audo AI provides two dedicated 48 kHz engines at $0.05/min ($1.88–$3.00/hr pay-as-you-go from $5, first minute free 3x/day with no sign-up). On noisy mics without heavy room echo, non-generative Audo AI Remove noise at 80% strength cut UNIVERSE-100 word errors by 2.76 percentage points below raw input (8.86% WER and 0.85% dropped words on non-reverb clips; 0.38% WER, 0.994 speaker similarity, and 18.5 kHz bandwidth on 48 kHz noise). On echoey rooms and muffled microphones, generative Audo AI Audio enhancer ranked #1 on reverberant speech (DNSMOS P.835 OVRL 3.35 on 48 kHz room echo, 17.8 kHz restored bandwidth on UNIVERSE-100, 35x real-time speed, and bit-identical MP4 video stream-copy).

Master UNIVERSE-100 scorecard (100 clips, 8.4 min speech, CC BY 4.0) paired with production speed, video handling, and effective hourly cost. Acoustic & speaker-embedding scores match within ±0.004 across shared anchors; Δ WER measures net percentage-point change vs. unprocessed noisy input under each run's pure CTC transcriber (checked November 18, 2026).
Model / ToolArchitectureWER (Δ vs. raw) ↓Dropped words ↓DNSMOS OVRL ↑PESQ / ESTOI ↑Speaker sim ↑Output bandwidth ↑Speed & effective $/hr
Audo AI — Remove noise (80% blend)48 kHz non-generative neural mask (noise_reduction=80)17.0% (−2.8 pp, HuBERT-CTC) · 8.9% on non-reverb4.2% (0.85% on non-reverb)2.51 (3.08 on 48k noise)1.73 / 0.759 (3.09 / 0.935 on 48k noise)0.929 (0.994 on 48k noise)7.6 kHz on 16k in · 18.5 kHz on 48k in33.7x RT (~33s / 20m, <2m / 1h); $1.88–$3.00/hr PAYG; bit-exact MP4 copy
Audo AI — Audio enhancer (100% default)48 kHz generative speech & dereverb restoration + −16 LUFS19.3% (−0.4 pp, HuBERT-CTC) · #1 on reverb (25.9%)6.9%3.16 (#1 on 48k reverb: 3.35, BAK 4.08)1.85 / 0.7870.938 (0.988 on 48k suite)17.8 kHz (#1 BWE from 16k in)35.2x RT (~38s / 20m, <2m / 1h); $1.88–$3.00/hr PAYG; bit-exact MP4 copy
Audo AI — Remove noise (100% default)48 kHz non-generative neural mask (full suppression)28.2% (+8.5 pp overall; 3.4% on pure additive noise)14.2% (over-gates heavy room echo tails)2.70 (3.28 on 48k noise, BAK 4.05)1.94 / 0.703 (#1 on 48k noise: 3.34 / 0.946)0.898 (#1 on 48k noise: 0.996)7.5 kHz on 16k in · 18.0 kHz on 48k in31.1x RT (~33s / 20m, <2m / 1h); $1.88–$3.00/hr PAYG; bit-exact MP4 copy
Cleanvoice Studio Sound 4.1 (published Nov 2026)Generative speech enhancer (processed per individual clip)15.5% (−1.5 pp, Omnilingual-CTC)3.0%3.351.88 / 0.8400.95517.3 kHz~5–10 min per 1h audio (6–12x RT); $1.50–$2.20/hr PAYG or $0.90–$1.10/hr sub; 3 GB cap
Adobe Podcast Enhance Speech v2 (published Nov 2026)Generative speech resynthesis (joined 8.4m UNIVERSE pass)17.1% (+0.1 pp, Omnilingual-CTC)6.8%3.082.07 / 0.8240.9459.5 kHzWeb app only (no API/MCP); Free: 30m/500MB audio · US$19.99/mo Firefly Pro: 2h/1GB
Auphonic Studio Voice Beta (published Nov 2026)Generative voice reconstruction + LUFS mastering (joined 8.4m pass)20.3% (+3.3 pp, Omnilingual-CTC)6.2%3.211.60 / 0.7930.94916.5 kHzAsync batch + REST API; 2h/mo free (with jingle) · $1.71–$2.40/hr one-time or $1.22–$1.44/hr sub
Dolby UNIVERSE (paper reference outputs)Score-based diffusion speech enhancer (16 kHz research model)21.0% (+1.3 pp, HuBERT-CTC · 22.3% Omnilingual-CTC)5.0% (5.1% Omnilingual)3.222.64 / 0.8630.9537.5 kHz (8.0 kHz Nyquist ceiling)Research checkpoint (Dolby.io commercial media APIs retired)
DeepFilterNet3 (open-source 48 kHz)Two-stage ERB + complex deep filter (MIT / Apache-2.0)20.6% (+0.9 pp, HuBERT-CTC · 22.6% Omnilingual-CTC)4.4% (9.6% Omnilingual)2.86 (3.39 on 48k noise)1.87 / 0.7630.9397.4 kHz on 16k in · 17.7 kHz on 48k noise~43x RT local CPU on short WAVs; $0 self-hosted WAV CLI (no MP4, dereverb, or LUFS)
ElevenLabs Voice Isolator (published Nov 2026)Neural speech isolator (joined 8.4m UNIVERSE pass, no slider)23.8% (+6.8 pp, Omnilingual-CTC)7.1%3.111.78 / 0.7540.92811.8 kHzREST API + MCP; 1,000 credits/min (~$9.90–$12.00/hr); hard 500 MB / 1h cap; API returns audio only
Descript Studio Sound (published Nov 2026)Generative voice regeneration inside timeline editor (joined 8.4m pass)25.4% (+8.4 pp, Omnilingual-CTC)8.1%3.241.85 / 0.8030.93317.3 kHz$16–$24/mo Hobbyist (10 media hrs/mo); re-encodes video; API cannot export media files directly
FFmpeg arnndn (RNNoise bd.rnnn)Lightweight GRU recurrent spectral mask (open-source CPU filter)21.9% (+2.2 pp, HuBERT-CTC)5.3%2.471.65 / 0.7070.9187.5 kHz on 16k in · 16.8 kHz on 48k noise~41x RT local CPU; $0 CLI audio filter (clips fast consonant onsets in heavy chatter)
FFmpeg afftdn (nf=−25)Stationary Wiener FFT noise gate (open-source CPU filter)21.3% (+1.6 pp, HuBERT-CTC)4.2%2.331.71 / 0.7090.9207.4 kHz on 16k in · 11.9 kHz on 48k noise~88x RT local CPU; $0 CLI audio filter (muffles 48 kHz highs and leaves non-stationary noise)
Unprocessed noisy input (reference)Raw validation clips before enhancement (not ranked)19.7% HuBERT-CTC · 17.0% Omnilingual-CTC4.5% HuBERT · 5.9% Omnilingual2.27 (2.273 ours · 2.27 Cleanvoice)1.55 / 0.717 (1.55 / 0.714 Cleanvoice)0.916 (0.9159 ours · 0.916 Cleanvoice)7.6 kHz PSD roll-off (8.0 kHz Nyquist)Baseline reference across 100 clips (1,415–1,423 words)

Listening tests

Press play, then switch tracks mid-sentence to compare.

Street & transit noise (48 kHz mic)

Clean 48 kHz studio voice over 7.5 dB SNR city bus rumble and hiss. Audo AI Remove noise strips background noise while keeping 0.999 speaker similarity, PESQ 3.19, and full 20.7–20.8 kHz natural sibilance.

Original noisy recording0:00 / 0:00

Source: VoiceBank-DEMAND 48 kHz test set (speaker p257, DEMAND bus noise, Valentini-Botinhao et al., Edinburgh DataShare 10283/2791, CC BY 4.0). All tracks loudness-matched to −18.0 LUFS.

Why one model fails half of real recordings: background noise vs. room echo

Most AI audio benchmarks average noisy studio microphones, reverberant halls, and telephone clips into one headline score. When we stratified both the 100-clip 16 kHz Dolby UNIVERSE set (59 non-reverb clips vs. 41 reverberant room clips) and our 24-clip full-band 48 kHz benchmark (12 close-mic noise recordings vs. 12 reverberant hall and muffled phone recordings), the rankings flipped completely depending on the acoustic defect—explaining why Audo AI separates Remove noise from Audio enhancer.

On clean 48 kHz microphones contaminated by cafe chatter, street traffic, bus rumble, or office HVAC noise, non-generative neural masking wins on waveform fidelity and word accuracy. Audo AI's Remove noise at 100% achieved the #1 wideband PESQ (3.34 vs. 2.67 for generative resynthesis), #1 ESTOI intelligibility (0.946 vs. 0.854), and #1 WavLM speaker identity similarity (0.996), while dialing Remove noise to 80% achieved the lowest word error rate (0.38% vs. clean CTC, 0.00% dropped words) and preserved 18.5 kHz of real top-end sibilance. Conversely, on reverberant rooms (0.55–1.05 s RT60) and muffled 250–3,600 Hz phone speech, every non-generative filter failed to clear smeared reflections (ESTOI 0.34–0.44), whereas Audo AI's generative Audio enhancer lifted 48 kHz room-reverb DNSMOS OVRL from 2.03 to 3.35, doubled ESTOI to 0.671, and rebuilt phone-band frequencies from 5.8 kHz out to 13.9 kHz.

  • Best for close-mic background noise (podcast mics, lavaliers, street & office noise): Audo AI Remove noise (80%–100%)

    On the 59 non-reverberant UNIVERSE clips, Audo AI Remove noise at 80% cut word errors to 8.86% with just 0.85% dropped words (beating raw input at 11.53%, Dolby UNIVERSE at 12.14%, DeepFilterNet3 at 13.47%, and generative Audio enhancer at 14.56%), without resynthesizing the speaker's vocal formants.

  • Best for room echo, distant camera mics, AirPods, and phone calls: Audo AI Audio enhancer (100%)

    On the 41 reverberant UNIVERSE clips, Audo AI Audio enhancer ranked #1 across all tested models in both word error rate (25.89%, cutting pure-reverb WER from 38.6% down to 14.9%) and perceptual quality (DNSMOS OVRL 3.18, ahead of Dolby UNIVERSE at 3.12 and DeepFilterNet3 at 2.74)—and also powers Audo's dedicated Remove echo workflow.

Full-band 48 kHz Creator Defect Split (24 clips: 12 VoiceBank-DEMAND 48 kHz close-mic noise clips vs. 6 reverberant room clips + 6 muffled phone-band clips, CC BY 4.0, measured November 18, 2026).
Model (48 kHz native)Close-mic noise: PESQ / ESTOI ↑Close-mic noise: Speaker sim & WERRoom echo (0.55–1.05s RT60): DNSMOS OVRL / ESTOI ↑Muffled phone (3.6 kHz): DNSMOS OVRL & Bandwidth ↑All 24 clips: DNSMOS OVRL / Speaker sim ↑
Audo AI — Audio enhancer (100%)2.67 / 0.8540.992 sim · 0.76% WER (0.0% dropped)3.35 OVRL (BAK 4.08) / 0.671 ESTOI (#1)3.29 OVRL (BAK 4.06) · 13.9 kHz rebuilt (#1)3.35 OVRL (#1) / 0.988 sim (#1)
Audo AI — Remove noise (100% mask)3.34 (#1) / 0.946 (#1)0.996 sim (#1) · 0.76% WER (0.0% dropped)2.27 OVRL / 0.345 ESTOI (echo stays)2.23 OVRL · 7.9 kHz (no BWE)2.76 OVRL / 0.955 sim
Audo AI — Remove noise (80% blend)3.09 / 0.9350.994 sim · 0.38% WER (#1) · 18.5 kHz (#1)2.07 OVRL / 0.363 ESTOI (echo stays)2.32 OVRL · 7.7 kHz (no BWE)2.64 OVRL / 0.957 sim
DeepFilterNet3 (48 kHz open source)3.33 / 0.933 (DNSMOS OVRL 3.39)0.995 sim · 0.76% WER2.66 OVRL / 0.442 ESTOI2.52 OVRL · 4.2 kHz2.99 OVRL / 0.967 sim
FFmpeg arnndn (RNNoise)2.50 / 0.8950.991 sim · 1.51% WER2.19 OVRL / 0.338 ESTOI2.37 OVRL · 4.3 kHz2.68 OVRL / 0.951 sim
FFmpeg afftdn (nf=−25)2.47 / 0.879 (rolls off highs to 11.9 kHz)0.989 sim · 0.38% WER2.11 OVRL / 0.338 ESTOI2.41 OVRL · 3.9 kHz2.58 OVRL / 0.946 sim
Unprocessed 48 kHz input (reference)2.22 / 0.8920.992 sim · 0.76% WER · 20.9 kHz2.03 OVRL / 0.334 ESTOI2.21 OVRL · 5.8 kHz2.53 OVRL / 0.953 sim

Word error rates, dropped syllables, and why the 80% strength slider matters

A speech enhancer that silences background hiss while swallowing quiet consonants damages downstream listening and transcription. In Cleanvoice's November 2026 UNIVERSE-100 study (scored with Meta Omnilingual ASR CTC 7B v2 across 1,423 reference words), five of the seven commercial and research models made more word errors than the 17.0% unprocessed input—reaching 20.3% on Auphonic Studio Voice, 22.3% on Dolby UNIVERSE, 22.6% on DeepFilterNet3, 23.8% on ElevenLabs Voice Isolator, and 25.4% on Descript Studio Sound—while only Cleanvoice Studio Sound 4.1 (15.5%, −1.5 pp) cut errors below raw audio and Adobe Podcast Enhance Speech stayed flat (17.1%, +0.1 pp).

Evaluating the exact same 100 Dolby UNIVERSE recordings with open-weight HuBERT-Large pure CTC (facebook/hubert-large-ls960-ft, greedy CTC decoding with zero language-model auto-correction across 1,415 clean reference words, where unprocessed audio scored 19.72% WER) reproduced that exact failure mode on conventional models—and revealed why Audo AI exposes a 10–100% strength control on both free previews and paid API calls. Running a non-generative noise suppressor at 100% on heavy room echo treats smeared reverberant tails as noise and gates them aggressively (14.21% dropped words across all 100 clips, vs. just 3.4% total WER on pure additive noise). Pulling Audo AI Remove noise to 80% strength preserved consonant transitions across the entire 100-clip corpus, dropping overall WER to 16.96% (−2.76 pp below raw input, winning 33 clips against 15 losses vs. DeepFilterNet3 and 37 wins vs. 21 losses against Dolby UNIVERSE), while Audo AI Audio enhancer at 100% also beat raw input at 19.29% WER (−0.42 pp, winning 41 clips vs. 16 losses against DeepFilterNet3).

Detailed UNIVERSE-100 pure CTC word-error breakdown (facebook/hubert-large-ls960-ft, 100 clips) stratified into 59 non-reverb clips vs. 41 reverberant clips.
Directly tested modelAll 100 clips: WER (Δ vs. raw) ↓Dropped / Changed / Added %59 Non-reverb clips: WER ↓41 Reverb clips: WER ↓Zero-error clips (of 100) ↑Head-to-head vs. DeepFilterNet3 (W / T / L)
Audo AI — Remove noise (80%)16.96% (−2.76 pp, #1)4.17% / 12.01% / 0.78%8.86% (0.85% dropped, #1)28.05%45 / 100 (#1 tie)33 fewer errors / 52 tied / 15 more (+51 words saved)
Audo AI — Audio enhancer (100%)19.29% (−0.42 pp, #2)6.86% / 11.52% (#1) / 0.92%14.56%25.89% (#1; 14.9% on pure reverb)45 / 100 (#1 tie)41 fewer errors / 43 tied / 16 more (+18 words saved)
Unprocessed noisy input19.72% (0.00 pp baseline)4.52% / 14.28% / 0.92%11.53%31.13%44 / 10030 fewer / 51 tied / 19 more (+12 words)
DeepFilterNet3 (48 kHz)20.57% (+0.85 pp)4.38% / 15.19% / 0.99%13.47%30.46%38 / 100— (reference opponent)
Dolby UNIVERSE (paper outputs)20.99% (+1.27 pp)5.02% / 15.05% / 0.92%12.14%33.33%43 / 10031 fewer / 42 tied / 27 more (−6 words)
FFmpeg afftdn (nf=−25)21.27% (+1.55 pp)4.17% / 15.97% / 1.13%12.90%32.93%42 / 10024 fewer / 49 tied / 27 more (−10 words)
FFmpeg arnndn (RNNoise)21.91% (+2.19 pp)5.30% / 16.18% / 0.42%13.84%33.17%40 / 10025 fewer / 47 tied / 28 more (−19 words)
Audo AI — Remove noise (100%)28.20% (+8.48 pp)14.21% / 12.72% / 1.27%15.79% (3.4% on pure noise)45.52% (gates heavy reverb)39 / 10026 fewer / 44 tied / 30 more (−108 words on reverb)

Production throughput, bit-exact MP4 video passthrough, and cost per hour

Academic speech benchmarks rarely test 60-minute episodes or 1080p/4K MP4 containers, yet processing bottlenecks, upload caps, and lossy video re-encodes decide whether a tool works in real creator and developer pipelines. On live production calls to the Audo v1 REST API (plus 30-day PostHog production medians across more than 5,200 succeeded jobs), Audo AI processed the joined 9.2-minute UNIVERSE-100 WAV in 15.70 seconds on Audio enhancer (35.2x real-time) and 16.37 seconds on Remove noise at 80% (33.7x real-time), a 10-minute 48 kHz lossless FLAC in 15.20 to 16.92 seconds (35.5x–39.5x real-time), a 20-minute recording in ~33 to 38 seconds, and a full 1-hour recording in under 2 minutes (~1m 41s–1m 42s)—roughly 3x to 5x faster than Cleanvoice's stated 5 to 10 minutes per hour.

To verify video integrity mathematically, we ran a 20-second 1920×1080 30 fps H.264 + stereo AAC MP4 (17,374,765 elementary video stream bytes, 600 YUV420p frames) through both Audo AI Remove noise and Audo AI Audio enhancer keeping the original container format. Extracting the raw H.264 elementary video stream via ffmpeg produced the exact same SHA-256 digest (522f07974ade4b79dd02d89ccee488eaa3617819051e744604bcbdf3cfc6c46c) before and after both tools while normalizing the cleaned audio track from −21.02 LUFS to −16.07 and −16.17 LUFS (true peak ≤ −1.18 dBTP). Timeline editors such as Descript re-encode video on export, Adobe Podcast gates MP4/MOV behind US$19.99/mo Firefly Pro with a 2-hour / 1 GB ceiling, and ElevenLabs' /v1/audio-isolation endpoint caps uploads at 500 MB / 1 hour and returns audio bytes only.

Production speed, file limits, MP4 video stream handling, loudness accuracy, and API / MCP support (verified November 18, 2026).
Platform1-min / 10-min / 1-hr audio speedMax upload size & durationMP4 / MOV video handlingLoudness targetREST API & MCP serverFree tier & paid pricing
Audo AI (Remove noise & Audio enhancer)5.1–6.2s (1m) · 15.2–16.9s (10m) · <2 min (1h)10 GB (guest & signed in), up to 10 hoursDirect MP4, MOV, MKV, WebM, AVI in/out; bit-identical video stream copy (0 generation loss)−16.0 LUFS (±0.15 LU measured, ≤ −1.0 dBTP)Keyless Free API (/api/try/*), REST API (/api/v1/*), & hosted MCP (audo.ai/mcp)1st min free 3x/day, no sign-up · $0.05/min ($3.00/hr on $5 pack; $1.88–$2.40/hr on $10–$100 non-expiring packs)
Cleanvoice Studio Sound 4.1Async queue (~5 to 10 min per 1h audio)Up to 3 GB web/API (5 GB MCP), no duration capAccepts MP4, MOV, WebM (and MKV on API/MCP)Podcast loudness normalization optionREST API (Python/JS SDKs) & hosted MCP (api.cleanvoice.ai/mcp)20s keyless preview + 30m one-time signup trial · $1.50–$2.20/hr PAYG (2-yr expiry) or $0.90–$1.10/hr monthly
Adobe Podcast Enhance Speech v2Browser queue (~1–3 min per 30m clip)Free: 30m, 500 MB (1h/day) · Paid: 2h, 1 GB (4h/day)Audio-only on Free; MP4 & MOV unlocked on paid Firefly ProFixed internal leveling (no LUFS target selector)None (browser web app only)Free audio under 30m (Adobe login) · US$19.99/mo Firefly Pro subscription
Auphonic (Noise Reduction & Studio Voice)Async cloud post-production queueChunked web/API upload; 3-min minimum bill per fileAccepts video containers; 4K cutting on paid tiers−16, −14, −18, −23, or −24 LUFS + Adaptive LevelerREST API, CLI, webhooks, & n8n node (ToS bars rival benchmarking)2h/mo free (adds audio jingle) · $1.71–$2.40/hr one-time credits or $1.22–$1.44/hr monthly
Descript Studio SoundProject timeline background render1 GB Free · 10–50 GB Paid (advises splitting 6h+ files)Full video editor (re-encodes H.264/HEVC video on export; 720p watermark on Free)Export loudness setting inside timeline renderEarly-access Underlord API & MCP (direct media file export not yet supported)60 media min/mo Free · $24/mo ($16/mo annual) Hobbyist for 10 media hrs/mo
ElevenLabs Voice IsolatorLow-latency sync & streaming APIHard 500 MB and 1-hour cap per file on all plansWeb UI accepts video; /v1/audio-isolation API returns audio stream onlyNo integrated EBU R 128 LUFS normalizerREST + streaming API & ElevenLabs MCP package10 min/mo Free (non-commercial) · $5–$99/mo (~$9.90–$12.00/hr at 1,000 credits/min)

How we tested—and verified 4-decimal calibration parity

Every direct measurement in this Audo AI benchmark ran on permissively licensed public audio corpora with clean ground-truth studio references, allowing both non-intrusive neural MOS scoring and reference-based signal and word-error verification. Track 1 uses the official 100-clip Dolby UNIVERSE validation archive (universe-validation_set-100.zip, Serrà et al., CC BY 4.0: 453.0 seconds of 16 kHz noisy and reverberant speech across VCTK and Harvard Sentences). Following the multi-vendor protocol used in Cleanvoice's November 2026 study, Audo AI processed the 100 clips joined into one lossless WAV with 1.0-second zero-padded guard intervals (552.04 seconds) at default loudness normalization (returning 48 kHz PCM_16 WAVs) and sliced each clip back at exact 3.0x sample offsets. Track 2 uses 24 true 48 kHz recordings (137.9 seconds, 12 female + 12 male utterances) from the Edinburgh DataShare 10283/2791 VoiceBank-DEMAND 48 kHz test set and CSTR VCTK 0.92 (close-mic channel 1, FFT-verified for real 14–22 kHz speech energy), split evenly into 12 close-mic noise clips, 6 pyroomacoustics reverberant rooms (0.55–1.05 s RT60, direct-path aligned to sample 0), and 6 muffled 250–3,600 Hz phone-band recordings.

All outputs were scored automatically in Python 3.12 with official Microsoft ITU-T P.835 DNSMOS ONNX (sig_bak_ovr.onnx with official polynomial calibration), wideband ITU-T P.862.2 PESQ (16 kHz, sub-50 ms cross-correlation lag compensated), ESTOI (pystoi), WavLM-Base+-SV (microsoft/wavlm-base-plus-sv) speaker-verification embedding cosine similarity against the unpadded clean reference, active-speech −50 dB power-spectral-density upper roll-off bandwidth (kHz), ITU-R BS.1770-4 integrated LUFS with 4x oversampled true peak, and HuBERT-Large (facebook/hubert-large-ls960-ft) greedy CTC transcription normalized with jiwer against both the clean-reference CTC transcript and ground-truth text. Comparing our scorer against Cleanvoice's published November 2026 numbers on all three shared anchor rows (Unprocessed input, Dolby UNIVERSE, and DeepFilterNet3) confirmed virtually identical acoustic calibration: 2.2729 vs. 2.27 (raw) and 3.2238 vs. 3.22 (Dolby) on DNSMOS OVRL, 2.6397 vs. 2.64 on PESQ-WB, 0.8631 vs. 0.863 on ESTOI, and 0.9159 vs. 0.916 on speaker similarity. Earlier 4-clip diagnostic numbers referenced in our other cleanup guides (Remove noise BAK 3.77–4.15, Audio enhancer BAK 4.20–4.23 and OVRL 3.42–3.52) come from a separate 4-recording 48 kHz field spot-check level-matched to −26 dBFS RMS rather than the heavier 16 kHz UNIVERSE validation mix.

Open benchmark dataset & raw scores

Audo AI Speech Enhancement & Noise Removal Benchmark Dataset (2026)

Per-clip and aggregate objective speech enhancement scores across 100 Dolby UNIVERSE validation clips (16 kHz, 800 model-clip evaluations) and 24 full-band 48 kHz VoiceBank-DEMAND / VCTK defect-split clips (168 model-clip evaluations), plus 1080p MP4 bitstream passthrough and production latency timings.

Variables measured
Word Error Rate (WER %), Dropped Words (%), Changed Words (%) · DNSMOS P.835 Overall (OVRL), Signal (SIG), and Background (BAK) · Wideband PESQ (1.0–4.5) and Extended Short-Time Objective Intelligibility (ESTOI) · Speaker Embedding Cosine Similarity (0.0–1.0) · Output Speech Bandwidth (kHz) and Processing Real-Time Factor (RTF)
Measurement techniques
ITU-T P.835 DNSMOS (Microsoft sig_bak_ovr.onnx) · Pure CTC Word Error Rate (facebook/hubert-large-ls960-ft + jiwer) · Speaker Verification Cosine Similarity (microsoft/wavlm-base-plus-sv) · ITU-T P.862.2 Wideband PESQ & Extended STOI (ESTOI) · Active-Speech −50 dB PSD Bandwidth Roll-Off (kHz) · ITU-R BS.1770-4 Integrated LUFS & 4x Oversampled True Peak (dBTP)
License & coverage
CC BY 4.0 · 2026-11

Where other tools scored higher & benchmark limitations

No vendor-run benchmark is independent—Audo AI makes Remove noise and Audio enhancer—and different models win on specific metrics, workloads, and pricing tiers:

  • 16 kHz aggregate studio MOS & high-volume monthly plans: Cleanvoice Studio Sound 4.1

    On Cleanvoice's November 2026 UNIVERSE-100 run (processed clip-by-clip rather than as one joined 8.4-minute file), Cleanvoice Studio Sound 4.1 posted a higher 16 kHz aggregate DNSMOS OVRL (3.35 vs. 3.16 on Audo AI Audio enhancer) and speaker similarity (0.955 vs. 0.938), and its $11–$90/mo subscriptions drop effective pricing to $0.90–$1.10/hr if you process 10 to 100 hours every month (plus mouth-click and breath removal, which Audo AI does not target).

  • Reference waveform alignment at 8 kHz Nyquist: Dolby UNIVERSE & DeepFilterNet3

    Dolby's 16 kHz UNIVERSE paper outputs scored highest on 16 kHz intrusive PESQ-WB (2.64) and ESTOI (0.863) because they reconstruct in-band 8 kHz speech without synthesizing new upper harmonics, and open-source DeepFilterNet3 ($0, MIT/Apache-2.0) edged Audo AI Remove noise slightly on no-reference 48 kHz noise-only DNSMOS OVRL (3.39 vs. 3.28) if you only process local WAV files and do not need room dereverb, MP4 passthrough, or −16 LUFS mastering.

  • Free short audio clips, free monthly podcast hours, or surgical DAW repair

    Adobe Podcast Enhance Speech cleans audio-only clips up to 30 minutes (1 hour/day) for $0; Auphonic includes 2 free hours a month of podcast leveling (with an audio jingle) and multitrack adaptive ducking; and iZotope RX 12 Standard ($399 one-time) lets audio engineers paint out individual chair squeaks or sirens by hand on a desktop spectrogram. Note also that Auphonic's terms restrict rival benchmarking, so Auphonic's row cites published third-party UNIVERSE-100 data rather than new automated API runs.

  • Dataset scope: single-speaker English VCTK utterances vs. multi-speaker crosstalk

    Both UNIVERSE-100 and VoiceBank-DEMAND use single-speaker English sentences. Overlapping two-host podcast crosstalk, tonal languages, live music beds (better served by Voice isolator or Remove music from video), and sudden transient noises like dog barks or keyboard clacks (better served by Sound remover) behave differently—always run the free 1-minute preview on your own file before paying.

Benchmark both Audo AI engines on your own recording for free

Drop any noisy or echoey audio or MP4/MOV video up to 10 GB to audition Audio enhancer (for room echo and muffled mics) or Remove noise at 80%–100% strength (to lock vocal identity)—first minute free, 3 times a day, with no sign-up.

Questions

Which AI speech enhancer keeps the most words intact without hallucinating?

It depends on whether your recording has room echo. On noisy recordings without room echo, non-generative Audo AI Remove noise at 80% strength achieved the lowest word error rate in our test—cutting errors on the 100-clip Dolby UNIVERSE set by 2.76 percentage points below the raw input (16.96% overall, 8.86% with 0.85% dropped words on the 59 non-reverb clips, and 0.38% WER on 48 kHz close-mic noise). Among generative speech restorers that also remove room echo, Cleanvoice Studio Sound 4.1 (15.5% vs. 17.0% raw, −1.5 pp) and Audo AI Audio enhancer (19.29% vs. 19.72% raw, −0.42 pp, and #1 at 25.89% on the 41 reverberant clips) were the only two generative models that reduced word errors instead of increasing them.

Should I use Audo AI Remove noise or Audio enhancer for my podcast or video?

Use Audo AI Remove noise (at 80%–100% strength) when the speaker's microphone already sounds clear and you want to strip background hiss, fan hum, cafe chatter, wind, or street traffic while keeping 0.994–0.996 speaker identity similarity and real 48 kHz vocal formants. Switch to Audo AI Audio enhancer (or Remove echo) when the voice echoes in a bare room, sounds distant on a camera or laptop mic, or is muffled like a phone or AirPods call—where Audio enhancer lifted 48 kHz room-reverb DNSMOS OVRL from 2.03 to 3.35 and rebuilt 3.6 kHz muffled speech out to 13.9 kHz. Both cost $0.05/min ($1.88–$3.00/hr pay-as-you-go from $5) and preview the first minute free 3 times a day.

How does Audo AI compare with Cleanvoice Studio Sound 4.1 and Adobe Podcast Enhance?

On 16 kHz UNIVERSE-100 inputs, Audo AI Audio enhancer restored wider upper speech bandwidth (17.8 kHz median vs. 17.3 kHz on Cleanvoice 4.1 and 9.5 kHz on Adobe Podcast Enhance Speech), processed audio roughly 3x to 5x faster than Cleanvoice (<2 minutes per hour vs. 5–10 minutes per hour), and accepts 10-hour / 10 GB audio and MP4/MOV video files via browser, REST API, and MCP (whereas Adobe has no API/MCP and caps free audio at 30 minutes / 500 MB and paid Firefly Pro at 2 hours / 1 GB). Cleanvoice 4.1 scored slightly higher on aggregate 16 kHz studio DNSMOS OVRL (3.35 vs. 3.16) and offers cheaper $0.90–$1.10/hr monthly plans above 10 hours/month, while Audo AI adds a separate non-generative Remove noise engine for zero-resynthesis vocal preservation.

Why did Audo AI's scorer match Cleanvoice's UNIVERSE-100 acoustic numbers to three decimal places?

Both Audo AI and Cleanvoice ran on the exact same 100 public 16 kHz clips from Dolby's UNIVERSE validation set (Serrà et al., CC BY 4.0) using Microsoft's official ITU-T P.835 DNSMOS ONNX model, wideband ITU-T P.862.2 PESQ, ESTOI, and WavLM speaker-verification embeddings. On the deterministic Unprocessed input and Dolby UNIVERSE paper outputs, Audo AI's evaluation pipeline produced 2.2729 vs. 2.27 and 3.2238 vs. 3.22 on DNSMOS OVRL, 2.6397 vs. 2.64 on PESQ-WB, 0.8631 vs. 0.863 on ESTOI, and 0.9159 vs. 0.916 on speaker similarity. Word error rates used open-weight HuBERT-Large pure CTC in Audo AI's direct runs vs. Meta Omnilingual ASR CTC 7B v2 in Cleanvoice's published table, so our master table reports both raw WER and net Δ WER vs. unprocessed input.

Does Audo AI re-encode MP4 or MOV video files during speech enhancement?

No. When you upload an MP4, MOV, MKV, WebM, or AVI file with default output settings, both Audo AI Remove noise and Audo AI Audio enhancer clean and normalize the audio track to −16 LUFS while copying the original H.264, HEVC, or ProRes video packets bit-for-bit without re-encoding. In our 1080p H.264 verification run, all 17,374,765 elementary video stream bytes and all 600 decoded video frames produced the exact same SHA-256 hash (522f0797...) before and after enhancement, avoiding the generation loss and slow render queues of timeline video editors.

Can I download the raw per-clip CSV scores or benchmark Audo AI from Python or MCP?

Yes. Every per-clip score across all 800 UNIVERSE-100 evaluations and 168 full-band 48 kHz evaluations is downloadable directly from this page as CC BY 4.0 CSV and JSON files. Developers and AI agents can reproduce or extend the Audo AI runs immediately via the Free API (POST /api/try/remove-noise and /api/try/enhance-voice), the v1 REST API (/api/v1/files and /api/v1/jobs), or the hosted MCP server at https://audo.ai/mcp.