Universal-3.6 Pro Realtime: trained on real conversations, built for real voice agents
Voice agents don’t fail on the audio most benchmarks are built from. They fail on short user replies, a phone number read out in chunks and chatter in the background. We built Universal-3.6 Pro Realtime to fix all three.
Voice agents pose a different challenge for a speech-to-text model than dictation or meeting transcription: turns are short, a one-word answer can decide the next action, and users are often on a noisy phone line with other voices nearby. Universal-3.6 Pro Realtime is trained for that setting, on tens of thousands of hours of real voice-agent and call-center conversations. On an in-house English voice-agent benchmark of 12,460 scripted caller scenarios voiced by real speakers, it makes 45% fewer errors on short replies than Universal-3.5 Pro Realtime, transcribes 29% less speech from people who were not talking to the agent, and switches to the wrong language in 74% fewer sessions. Against 20 other streaming systems on the same benchmark, it has the lowest word and entity error rates, 24% and 21% below the next system. Beyond English, it adds 14 languages, including Korean and Russian, and has the lowest code-switching error of all streaming systems on every language pair. Voice focus for a competing talker and entity-aware endpointing complete the release.
1Introduction
Universal-3.5 Pro Realtime launched in June 2026, and within weeks it was running inside voice agents, contact-center software and meeting assistants. Over the following three months the teams building these products told us where it fell short. Their reports were consistent, and they described failures that average word error rate on public benchmarks barely registers:
- The wrong language. The most frequent report: accented English, most often Indian-accented, written out in another language’s script, and short or noisy turns returned in a language nobody spoke.
- Turns cut short. A turn ended while the user was still speaking, often in the middle of a phone number, an account number or an amount.
- Lost short replies. One-word answers such as “yes” or “no” dropped or misrecognized, although they decide what the agent does next.
- Other people’s speech. A television or a person at the next desk transcribed as the user; in the hardest cases, a second person talking right beside them.
- Missing languages. Korean and Russian above all, and better handling of speakers who mix English with another language.
The first four come from properties of real calls that standard speech corpora lack: short replies, pauses in the middle of a thought, digits spoken in chunks, accents, telephone channels and other voices in the room. Universal-3.6 Pro Realtime addresses them as follows:
- Training on real voice-agent and call-center conversations, for the wrong-language, short-reply and background-speech failures (Section 2).
- Training on incomplete sentences, so that a pause is not read as the end of a turn, and entity-aware endpointing, which holds a turn open while a phone number or an ID is still being spoken (Section 3).
- A new voice-focus model, applied to the audio the model recognizes and not only to where turns start and stop, for a competing talker (Section 4).
- Fourteen new languages, including Korean and Russian, stricter per-language data quality control, and fewer code-switching errors (Section 5).
Sections 2 to 5 compare Universal-3.6 Pro Realtime with Universal-3.5 Pro Realtime, the model in production, on the same audio and at the same latency operating point. Section 6 compares it with other streaming systems, using the runs behind our streaming benchmarks page.
2Voice-agent robustness
Voice-agent traffic stresses a speech recognizer in ways standard speech corpora rarely do. The reply that decides the next step is often one or two words, users call over narrowband telephone lines with every kind of accent, and other voices reach the microphone. The resulting failures are fundamental model issues that show up specifically in voice-agent traffic: turns written in the wrong language, short replies that go missing, and other people’s speech transcribed as the user’s, three of the problems customers reported (Section 1). The largest change in Universal-3.6 Pro is therefore in its training data: tens of thousands of hours of real voice-agent and call-center conversations, which contain one- and two-word confirmations, hesitations and restarts, accented speech, narrowband telephone channels, loud rooms, and voices from people who are not talking to the agent. It also deepens coverage of accented, conversational English, so an accent is less often read as a change of language. Figure 1 pairs four of these properties with the benchmark metric that measures each; the largest reductions fall on exactly these behaviors.
Table 1 compares Universal-3.5 Pro and Universal-3.6 Pro on our in-house English voice-agent benchmark: 12,460 scripted caller scenarios voiced by real speakers, recorded with no, low or heavy background noise in roughly equal parts. Word error rate is reported twice: the normalized rate compares transcripts after normalizing case, punctuation, numbers and abbreviations, while the raw rate also counts those formatting differences, which is why it is higher. Three further metrics are specific to voice agents: short response error rate, background-speech leakage and language confusion.
Table 1. English voice-agent benchmark, 12,460 scripted scenarios: Universal-3.5 Pro against Universal-3.6 Pro.
| metric | Universal-3.5 Pro | Universal-3.6 Pro | change |
|---|---|---|---|
| Short response error rate | 2.65% | 1.45% | −45% |
| Word error rate | 13.59% | 12.32% | −9% |
| Normalized word error rate | 5.80% | 5.13% | −12% |
| Language confusion (share of sessions) | 0.55% | 0.14% | −74% |
| Background-speech leakage (words per 1,000) | 3.53 | 2.52 | −29% |
Universal-3.6 Pro improves every metric. The largest gains are on the voice-agent metrics the new training data targets (short responses, language confusion and background-speech leakage); aggregate word error rate improves less.
2.1Short responses
Short responses such as “yes”, “no thank you” and “that’s it” carry the decision in a voice-agent turn: one misrecognized confirmation sends the call down the wrong path. The short response error rate covers the most common replies in voice-agent conversations and scores them by meaning. An automatic check counts a reply as an error only when the words change what was said, so “yeah” for “yes” or “thank you” for “thanks” is not an error, while a different response, a different word or a dropped response is. On this scoring Universal-3.5 Pro gets 2.65% of short responses wrong and Universal-3.6 Pro 1.45%, a 45% reduction.
2.2Background-speech leakage
Background-speech leakage counts the words that come from other people: every word the model adds inside a turn that overlaps no speech from the user, per 1,000 reference words. It is a transcript of the television, the coworker or the person at the next desk, delivered to the agent as if the user had said it. Universal-3.6 Pro reduces it by 29%, from 3.53 to 2.52 words per 1,000. Of the words Universal-3.6 Pro no longer inserts, 39% were in turns with no user speech; the rest were in the user’s own turns. These results are with voice focus, the optional filter described in Section 4, turned off: the model keeps chatter and televisions out of the transcript on its own.
2.3Language confusion
Universal-3.5 Pro and Universal-3.6 Pro both identify the language automatically. Universal-3.5 Pro occasionally committed a turn to the wrong language: an accented English speaker, a borrowed phrase or background chatter could switch a turn into another language’s script in the middle of a conversation. For a voice agent such a turn is unusable, and intent detection, entity capture and the language model’s reply fail with it. We count a session as affected if any of its output is in the wrong script. That rate falls from 0.55% of sessions to 0.14%, roughly one session in seven hundred. We attribute the change to the deeper coverage of accented, conversational English in the training data.
2.4Robustness to background noise
Figure 2 splits the benchmark by background noise (none, low, heavy). Every level improves on normalized word error rate and short response error rate, and the gain grows with the noise: normalized word error rate falls 7%, 9% and 16% respectively, and short response error rate 38%, 42% and 52%. Background-speech leakage falls 17% on clean audio, 26% under low and 31% under heavy noise.
3Turn detection
A voice agent has to decide, turn by turn, when the user has finished speaking; the part of the system that makes this decision is the endpointer. It has to make the call from a pause, and a pause is ambiguous: people stop in the middle of a sentence to think, and read numbers out in chunks. End too early and the user is cut off, often in the middle of a phone number or an amount, the failure customers reported (Section 1); end too late and every reply is delayed.
How turns are detected. Turn detection runs in the streaming stack around the recognition model, and that stack is the same for Universal-3.5 Pro and Universal-3.6 Pro (Figure 3). A voice-activity detector segments the incoming audio into speech and pauses, and the recognition model transcribes each turn as it is spoken. When the user pauses for 128 ms, the endpointer asks the model itself whether the turn is over: the model transcribes the turn so far, and if the transcript reads as a finished sentence, the turn ends there. If it is unfinished, or ends in a phone number or an ID that may still continue, the endpointer keeps listening, and the turn ends after 1,280 ms of silence. The quick exit keeps replies fast, but it is also where turns get cut short: a model that writes an unfinished thought as a finished sentence ends the turn with it.
Incomplete sentences. Users often pause before they have finished a thought, and that pause is where the quick exit can end a turn too early. Universal-3.6 Pro was trained on such incomplete sentences from real conversations.
We measure end-of-turn decisions on the voice-agent benchmark (Section 2). Each end-of-turn event the model emits is matched against the reference end of turn: precision is the share of events that land at a real end of turn rather than in the middle of one, recall is the share of real turn ends that receive an event, and latency is the time from the true end of speech to the event.
Table 2. End-of-turn decision quality on the voice-agent benchmark, 12,460 scripted scenarios. pp: percentage points.
| endpoint decision quality | Universal-3.5 Pro | Universal-3.6 Pro | change |
|---|---|---|---|
| End-of-turn F1 | 82.2% | 84.7% | +2.4 pp |
| End-of-turn precision | 74.1% | 77.9% | +3.8 pp |
| End-of-turn recall | 92.4% | 92.8% | held |
| End-of-turn latency, median | 537 ms | 537 ms | held |
| End-of-turn latency, 90th percentile | 1,593 ms | 1,606 ms | +13 ms |
Table 2 compares the end-of-turn decisions of Universal-3.5 Pro and Universal-3.6 Pro. Because the stack is the same, the differences come from the recognition model. Universal-3.6 Pro raises precision by 3.8 percentage points, from 74.1% to 77.9%, with recall held (92.4% against 92.8%), and F1 by 2.4 points. Median end-of-turn latency is unchanged at 537 ms and the 90th percentile moves by 13 ms, so the accuracy comparisons are made at the same latency.
Trained on incomplete sentences from real conversations, Universal-3.6 Pro is much better at recognizing from the acoustic evidence that what the user has said so far is not yet semantically complete, so fewer turns are ended while the user is still mid-thought.
3.1Entity-aware endpointing
Digit strings, account identifiers and spelled-out names break the cues an endpointer relies on: they arrive in chunks separated by real pauses, have no syntactic ending, and are read slowly on purpose. A conventional endpointer splits one phone number across several turns, and every downstream consumer (the language model, validation logic, a payment call) receives fragments.
Entity-aware endpointing keeps an entity in one turn (Figure 4). If, at the 128 ms check, the transcript ends like an entity still in progress (a phone number, an account ID, an email, an address or a spelled-out name), the turn is held open until the 1,280 ms silence limit, even when the model has already written a sentence ending. The hold is spent on roughly one turn in ten.
Table 3. What entity-aware endpointing buys: Universal-3.6 Pro Realtime with the entity hold on and off, on the same 12,460 scenarios.
| metric | entity hold off | entity hold on | change, on vs off |
|---|---|---|---|
| End-of-turn latency, median | 536 ms | 536 ms | 0% |
| End-of-turn latency, 90th percentile | 829 ms | 1,643 ms | +98% |
| End-of-turn precision | 72.2% | 77.8% | +5.7 pp |
| End-of-turn recall | 92.9% | 92.8% | held |
| Normalized word error rate | 6.01% | 5.16% | −14.2% |
| Phone numbers, entity error rate | 7.02% | 2.20% | −68.7% |
| Codes & identifiers, entity error rate | 13.76% | 9.73% | −29.3% |
| Email addresses, entity error rate | 21.97% | 18.27% | −16.8% |
| Street addresses, entity error rate | 30.91% | 27.16% | −12.1% |
Table 3 compares Universal-3.6 Pro Realtime with the entity hold on and off. The hold costs only in the slow tail: median latency is identical, and the 90th percentile moves from 0.83 s to 1.64 s. In exchange the fast exit no longer fires inside entities. The entity error rate, the share of entities with any word transcribed wrongly, falls from 7.0% to 2.2% for phone numbers, less than a third of its value without the hold, and phone-number character error falls from 14.5% to 3.1%. Entity-aware endpointing is on by default and also works with Universal-3.5 Pro Realtime.
4Voice focus
A voice agent should transcribe only the person talking to it, yet real calls often carry other voices, and the model has to decide whose speech belongs in the transcript. Section 2 showed that Universal-3.6 Pro makes this call well for background speech, such as a television or chatter in the room, which it keeps out of the transcript without any filtering. The hard case, and one customers reported (Section 1), is a competing talker: a second person close to the microphone, such as a coworker or a family member, whose speech is as clear as the user’s, so the model transcribes both. Voice focus is an optional filter for that case; it is off by default and should be turned on only when a competing talker is present. Voice focus changes in two ways in Universal-3.6 Pro. It runs a new voice-focus model, and its filtered audio now also reaches the recognition model; in Universal-3.5 Pro the filtered audio only decided where the user’s turns start and stop (Figure 5, top).
Figure 5 (bottom) compares Universal-3.5 Pro and Universal-3.6 Pro with voice focus off and on, at its default setting, on five test sets from three sources, none of them seen in training, grouped by whether a competing talker is present. In every set the reference covers only the main speaker, so each word transcribed from anyone else counts as an error.
- With a competing talker
- ai-coustics test calls (
aic_test_calls_en): clips cut from real calls recorded over meeting and telephony platforms, with other people talking in the background and human-verified transcripts of the main speaker. - Krisp Voice Isolation Benchmark, Work and Call Center (Krisp): recordings from home offices, meeting rooms and open-plan offices, and from live call centers, made in real rooms with no synthetic mixing. A second person speaks near the microphone, both over the main speaker and alone.
- ai-coustics test calls (
- Without a competing talker
- Krisp Voice Isolation Benchmark, Phone Calls: real mobile calls, most of them from cars, with road, street and crowd noise over band-limited telephone audio and no deliberate second speaker.
- Heavy noise: recordings from the heavy-background-noise condition of our in-house English voice-agent benchmark (Section 2), with one speaker at the microphone.
With a competing talker, turning voice focus on in Universal-3.6 Pro cuts the error rate by two thirds or more. On the ai-coustics test calls Universal-3.6 Pro reaches 20.2% word error rate against 53.8% for Universal-3.5 Pro, a 62% reduction. With voice focus off the two models are within two points, so the gain comes from what the model hears, not from the model alone. On the Krisp recordings the picture repeats: with voice focus on, Universal-3.6 Pro’s error rate falls from 34.8% to 6.9% on Work and from 19.7% to 4.6% on Call Center, 76% over the whole Krisp set. What disappears is the other person: transcribed second-talker words fall from 18.9 to 1.8 per 100 reference words.
Without a competing talker there is a small cost, which is why the feature is off by default. On heavy noise Universal-3.6 Pro goes from 8.4% to 9.2%; on clean single-speaker recordings from the same benchmark the cost is 0.2 points; on Krisp Phone Calls it is 2.2 points, all of it deleted user speech. Turn voice focus on when other people can be heard near the microphone, and leave it off otherwise. Its strength is configurable, so an application can trade how much of other people’s speech is suppressed against how much of the user’s own speech is kept; our documentation describes the setting.
5Multilingual recognition
A single realtime model serving many languages has two jobs: recognize each language well, and tell the languages apart without being told which one is spoken, including when a speaker switches between English and another language mid-conversation. Coverage matters as much as accuracy: Korean and Russian were the languages customers asked for most (Section 1). Multilingual training data was expanded for every supported language, including low-resource ones, and paired with strict per-language quality control: audio whose spoken language did not match its label was corrected or dropped.
Universal-3.6 Pro serves every language from one model with automatic language identification; no per-language endpoint or language hint is required. For Chinese, Japanese, Korean and Cantonese we report character error rate (CER) instead of word error rate, since these languages are not written with spaces between words. Figure 6 shows, for each of the 17 previously supported non-English languages, the relative change in error rate on Common Voice and FLEURS. Common Voice error falls in 16 of 17 languages, by 20% on average, and FLEURS error in all 17, by 11% on average. The two sets can disagree: Hindi’s Common Voice error rises 19% while its FLEURS error falls 39%. Code-switching errors, on 20 public sets that mix English with German, Spanish, French, Italian or Portuguese, fall 16% on average; 19 of 20 improve and none regress. By language pair the reduction ranges from 14% for English–Spanish and English–French to 20% for English–Portuguese.
5.1Fourteen new languages
Table 4 gives error rates for the 14 languages added in this release. Universal-3.5 Pro does not support them, so there is no baseline to compare with. Norwegian Bokmål was already supported; Norwegian Nynorsk is new.
Table 4. Languages added in Universal-3.6 Pro. FLEURS, normalized WER, *CER for Korean and Cantonese; automatic language detection, no language hints.
| language | error rate | language | error rate |
|---|---|---|---|
| Korean* | 4.5% | Afrikaans | 20.5% |
| Catalan | 5.3% | Marathi | 20.7% |
| Galician | 5.9% | Zulu | 27.1% |
| Russian | 7.4% | Urdu | 27.4% |
| Romanian | 15.8% | Xhosa | 29.1% |
| Persian | 16.2% | Cantonese* | 29.5% |
| Estonian | 17.6% | Norwegian Nynorsk† | 35.4% |
†Nynorsk has no FLEURS set; measured on short-form read speech instead.
6Competitive benchmarking
On English voice-agent audio, Universal-3.6 Pro Realtime has the lowest word and entity error rates of 21 streaming systems, and it has the lowest code-switching error on every language pair. On multilingual read speech it is third, behind Google Gemini 3.5 Live and Soniox Realtime on FLEURS and behind Speechmatics Enhanced and AWS Transcribe on Common Voice.
We compared Universal-3.6 Pro Realtime with the other streaming systems on our streaming benchmarks page, on the in-house English voice-agent benchmark and on public multilingual and code-switching test sets. The voice-agent benchmark is our own, built around the failures customers reported; the multilingual and code-switching sets are public. Every system is scored with the same normalizer and metrics. Universal-3.6 Pro was evaluated again for this comparison through our production API, so its values differ from the earlier tables. In the multilingual comparison every system, ours and the competitors’, is told the language of each test set in the form its API accepts. Code-switched speech is transcribed without a language hint.
Table 5. English voice-agent benchmark: Universal-3.6 Pro against the five other systems with the lowest normalized word error rate, Deepgram Nova-3, Deepgram’s most accurate system on this benchmark, and Zoom Scribe (Live), whose entity error rate is level with Soniox Realtime’s. The first row averages entity error rate over all entity classes; the next ten give single classes. Normalized word error rate is the mean over the three noise levels unless the row names one. The last two columns give Universal-3.6 Pro’s rank among all systems and its relative difference from the best other system in the row, including systems not shown.
| metric | Universal-3.6 Pro | Soniox Realtime | Meta Muse Voice Transcribe | ElevenLabs Scribe v2 | Speechmatics Enhanced | Google Gemini 3.5 Live | Deepgram Nova-3 | Zoom Scribe (Live) | Rank | vs closest |
|---|---|---|---|---|---|---|---|---|---|---|
| Entity error rate | ||||||||||
| All entities | 14.41% | 18.19% | 22.22% | 18.50% | 20.60% | 20.13% | 26.10% | 18.20% | 1 of 21 | −21% |
| Names | 10.95% | 16.57% | 23.85% | 14.85% | 23.84% | 15.78% | 24.25% | 16.48% | 1 of 21 | −26% |
| Phone numbers | 2.39% | 2.73% | 10.00% | 3.35% | 3.21% | 7.31% | 4.48% | 9.55% | 1 of 21 | −12% |
| Email addresses | 17.22% | 22.73% | 20.04% | 21.24% | 24.18% | 23.92% | 34.42% | 22.42% | 1 of 21 | −14% |
| Street addresses | 28.98% | 27.85% | 38.45% | 28.62% | 31.63% | 32.38% | 30.37% | 29.99% | 4 of 21 | +8% |
| Dates and times | 18.65% | 30.76% | 20.03% | 26.51% | 31.12% | 30.42% | 43.71% | 20.93% | 1 of 21 | −6% |
| Numbers and currency | 12.95% | 19.69% | 17.44% | 17.18% | 17.11% | 17.28% | 16.75% | 14.94% | 1 of 21 | −13% |
| Codes and identifiers | 10.01% | 8.17% | 27.82% | 12.38% | 19.45% | 16.90% | 27.65% | 13.95% | 2 of 21 | +23% |
| Places | 5.95% | 7.73% | 8.44% | 8.93% | 6.23% | 9.35% | 10.99% | 7.05% | 1 of 21 | −4% |
| Medical terms | 16.09% | 19.11% | 25.60% | 24.71% | 17.42% | 20.46% | 29.05% | 20.22% | 1 of 21 | −8% |
| Technical terms | 20.91% | 26.52% | 30.49% | 27.22% | 31.78% | 27.45% | 39.38% | 26.35% | 1 of 21 | −21% |
| Normalized word error rate | ||||||||||
| Mean over noise levels | 5.19% | 6.83% | 7.48% | 7.78% | 8.14% | 8.48% | 8.64% | 9.36% | 1 of 21 | −24% |
| No background noise | 4.37% | 5.98% | 6.92% | 6.18% | 6.96% | 7.44% | 7.53% | 8.04% | 1 of 21 | −27% |
| Low background noise | 5.08% | 6.84% | 7.50% | 7.53% | 7.96% | 8.49% | 8.45% | 9.29% | 1 of 21 | −26% |
| Heavy background noise | 6.13% | 7.68% | 8.02% | 9.64% | 9.49% | 9.52% | 9.93% | 10.74% | 1 of 21 | −20% |
BestWorstCells are coloured by rank within each row (solid blue is best, white is worst). Lower is better in every row.
In Table 5, Universal-3.6 Pro has the lowest entity error rate, 21% below Soniox Realtime, the next lowest, and the lowest error on eight of the ten entity classes, from 4% lower than the next system on places to 26% lower on names. They include phone numbers and amounts, two of the details customers reported being split across turns (Section 1): 12% and 13% below the next system. It is second on codes and identifiers, 23% above Soniox Realtime, and fourth on street addresses, 4% above Soniox Realtime and 1% above ElevenLabs Scribe v2. Universal-3.6 Pro also has the lowest normalized word error rate at every noise level, 20% to 27% below Soniox Realtime. Against Deepgram Nova-3, Universal-3.6 Pro’s entity error rate is 45% lower and its normalized word error rate 40% lower.
Table 6. Normalized word error rate on the English voice-agent benchmark by the speaker’s accent or dialect, for the systems of Table 5. The last two columns give Universal-3.6 Pro’s rank among all systems and its relative difference from the best other system in the row, including systems not shown.
| metric | Universal-3.6 Pro | Soniox Realtime | Meta Muse Voice Transcribe | ElevenLabs Scribe v2 | Speechmatics Enhanced | Google Gemini 3.5 Live | Deepgram Nova-3 | Zoom Scribe (Live) | Rank | vs closest |
|---|---|---|---|---|---|---|---|---|---|---|
| Normalized word error rate | ||||||||||
| Native speakers, all | 4.68% | 6.37% | 6.81% | 7.10% | 7.62% | 7.94% | 7.90% | 8.54% | 1 of 21 | −27% |
| US, General American | 4.62% | 6.29% | 6.69% | 7.04% | 7.66% | 7.76% | 7.83% | 8.12% | 1 of 21 | −27% |
| US, South | 4.90% | 7.16% | 7.21% | 7.57% | 8.67% | 8.82% | 8.03% | 10.05% | 1 of 21 | −32% |
| US, Northeast | 3.95% | 5.67% | 6.33% | 6.42% | 6.98% | 7.08% | 7.27% | 6.85% | 1 of 21 | −30% |
| US, African American English | 4.76% | 6.79% | 7.12% | 7.41% | 7.78% | 8.25% | 7.81% | 8.94% | 1 of 21 | −30% |
| US, Asian American | 4.80% | 6.50% | 7.32% | 6.76% | 7.14% | 7.91% | 8.12% | 9.80% | 1 of 21 | −26% |
| US, other | 5.32% | 6.88% | 7.84% | 7.70% | 8.71% | 8.77% | 8.76% | 11.80% | 1 of 21 | −23% |
| UK | 5.74% | 6.83% | 7.49% | 9.04% | 8.13% | 10.52% | 9.36% | 11.04% | 1 of 21 | −16% |
| Canada | 4.04% | 5.22% | 6.31% | 5.42% | 6.34% | 6.29% | 6.81% | 5.82% | 1 of 21 | −23% |
| Native, other | 5.55% | 6.86% | 7.92% | 7.91% | 7.85% | 8.48% | 9.31% | 10.38% | 1 of 21 | −19% |
| Non-native speakers, all | 7.40% | 8.82% | 10.38% | 10.72% | 10.38% | 10.85% | 11.82% | 12.90% | 1 of 21 | −16% |
| Indian-accented | 6.64% | 7.78% | 9.49% | 8.90% | 8.74% | 8.14% | 10.23% | 12.49% | 1 of 21 | −15% |
| Spanish-accented | 6.51% | 8.33% | 9.70% | 9.88% | 9.82% | 10.83% | 10.94% | 11.70% | 1 of 21 | −22% |
| Non-native, other | 7.66% | 9.30% | 10.20% | 11.40% | 10.84% | 11.59% | 12.23% | 12.64% | 1 of 21 | −18% |
BestWorstCells are coloured by rank within each row (solid blue is best, white is worst). Lower is better in every row.
Table 6 breaks the benchmark down by the speaker’s accent or dialect. Universal-3.6 Pro has the lowest normalized word error rate in every group, for native and non-native speakers alike, 15% to 32% below Soniox Realtime, the next lowest in every group. The margin is smallest, 15%, for Indian-accented speakers, the accent behind the most frequent customer report in Section 1, and largest for speakers from the US South.
Table 7. Multilingual speech with the language specified: error rate per language on FLEURS and Common Voice, for Universal-3.6 Pro and six competing systems. Each mean is taken over all 18 languages, for the systems with results in all of them, and its rank counts only those systems. Normalized WER; *CER for Chinese, Japanese and Korean. The last two columns give Universal-3.6 Pro’s rank among all systems and its relative difference from the best other system in the row, including systems not shown.
| metric | Universal-3.6 Pro | Google Gemini 3.5 Live | Soniox Realtime | AWS Transcribe | ElevenLabs Scribe v2 | Meta Muse Voice Transcribe | Speechmatics Enhanced | Rank | vs closest |
|---|---|---|---|---|---|---|---|---|---|
| FLEURS, mean of 18 languages | 8.44% | 6.39% | 7.93% | 8.95% | 9.33% | 10.01% | 11.06% | 3 of 13 | +32% |
| Common Voice, mean of 18 languages | 7.59% | 9.16% | 9.50% | 7.46% | 14.59% | 10.24% | 7.35% | 3 of 13 | +3% |
BestWorstCells are coloured by rank within each row (solid blue is best, white is worst). Lower is better in every row.
Per-language results: 18 languages on FLEURS and Common Voice
| metric | Universal-3.6 Pro | Google Gemini 3.5 Live | Soniox Realtime | AWS Transcribe | ElevenLabs Scribe v2 | Meta Muse Voice Transcribe | Speechmatics Enhanced | Rank | vs closest |
|---|---|---|---|---|---|---|---|---|---|
| FLEURS | |||||||||
| Arabic | 12.06% | 6.65% | 7.16% | 8.54% | 10.84% | 8.87% | 6.53% | 11 of 15 | +85% |
| Chinese* | 13.96% | 9.42% | 10.31% | 12.30% | 10.12% | 13.79% | 33.76% | 7 of 15 | +48% |
| Danish | 12.66% | 9.13% | 11.42% | 11.47% | 12.47% | 13.63% | 11.77% | 7 of 13 | +39% |
| Dutch | 5.90% | 4.97% | 7.66% | 7.11% | 7.01% | 8.56% | 8.47% | 2 of 18 | +19% |
| Finnish | 10.82% | 7.47% | 7.96% | 9.80% | 9.52% | 10.73% | 7.98% | 9 of 13 | +45% |
| French | 5.01% | 5.36% | 7.76% | 7.64% | 6.89% | 7.98% | 6.55% | 1 of 19 | −7% |
| German | 4.99% | 4.65% | 6.87% | 7.99% | 5.84% | 6.44% | 6.85% | 2 of 19 | +7% |
| Hebrew | 18.21% | 18.89% | 19.18% | 22.56% | 26.96% | 26.96% | 18.48% | 1 of 13 | −1% |
| Hindi | 7.50% | 5.02% | 6.33% | 6.31% | 15.81% | 8.14% | 6.98% | 6 of 17 | +49% |
| Italian | 3.98% | 3.18% | 4.91% | 4.63% | 4.34% | 5.97% | 5.38% | 2 of 19 | +25% |
| Japanese* | 6.30% | 5.00% | 5.81% | 7.00% | 8.59% | 9.76% | 31.33% | 4 of 18 | +26% |
| Korean* | 4.35% | 3.77% | 4.38% | 4.90% | 4.58% | 6.14% | 5.38% | 2 of 15 | +15% |
| Portuguese | 5.48% | 4.74% | 6.23% | 8.15% | 6.39% | 8.71% | 7.30% | 2 of 19 | +16% |
| Russian | 7.46% | 5.32% | 7.28% | 9.40% | 6.62% | 10.14% | 10.01% | 4 of 17 | +40% |
| Spanish | 4.11% | 3.75% | 5.00% | 5.28% | 4.58% | 6.45% | 4.72% | 3 of 19 | +10% |
| Swedish | 6.60% | 7.28% | 10.26% | 10.01% | 10.95% | 10.99% | 9.89% | 1 of 14 | −9% |
| Turkish | 8.56% | 5.29% | 7.29% | 8.61% | 7.72% | 10.50% | 9.06% | 4 of 13 | +62% |
| Vietnamese | 13.88% | 5.22% | 6.99% | 9.33% | 8.78% | 6.48% | 8.57% | 9 of 13 | +166% |
| Common Voice | |||||||||
| Arabic | 9.42% | 9.59% | 10.22% | 9.57% | 20.62% | 9.32% | 6.07% | 4 of 15 | +55% |
| Chinese* | 9.15% | 13.28% | 14.11% | 8.52% | 12.14% | 16.37% | 25.07% | 2 of 15 | +7% |
| Danish | 9.42% | 11.49% | 9.53% | 7.30% | 14.50% | 14.59% | 7.27% | 4 of 13 | +30% |
| Dutch | 2.65% | 3.27% | 4.58% | 2.69% | 5.48% | 5.67% | 2.44% | 2 of 18 | +9% |
| Finnish | 9.26% | 11.49% | 7.69% | 5.68% | 14.55% | 13.62% | 3.38% | 5 of 13 | +174% |
| French | 6.86% | 9.34% | 12.17% | 8.80% | 10.24% | 10.44% | 7.18% | 1 of 19 | −4% |
| German | 3.58% | 5.19% | 7.58% | 5.84% | 7.13% | 6.60% | 3.67% | 1 of 19 | −2% |
| Hebrew | 12.35% | 17.83% | 12.51% | 14.98% | 43.38% | 18.37% | 9.39% | 3 of 13 | +32% |
| Hindi | 5.25% | 4.68% | 6.64% | 5.09% | 21.60% | 5.78% | 3.87% | 6 of 17 | +36% |
| Italian | 3.32% | 4.19% | 6.30% | 4.18% | 6.11% | 5.53% | 4.70% | 2 of 19 | +7% |
| Japanese* | 16.17% | 21.71% | 19.18% | 15.92% | 20.73% | 20.25% | 23.19% | 2 of 18 | +2% |
| Korean* | 6.77% | 7.51% | 8.72% | 7.50% | 11.04% | 9.63% | 7.16% | 1 of 15 | −5% |
| Portuguese | 4.88% | 5.77% | 7.42% | 4.93% | 9.71% | 6.39% | 5.18% | 2 of 19 | +13% |
| Russian | 8.07% | 6.45% | 9.00% | 4.81% | 10.19% | 7.15% | 7.07% | 7 of 17 | +68% |
| Spanish | 3.19% | 4.05% | 6.02% | 4.14% | 5.78% | 4.92% | 2.54% | 2 of 19 | +26% |
| Swedish | 4.78% | 9.73% | 8.59% | 5.88% | 12.89% | 11.06% | 4.00% | 2 of 14 | +20% |
| Turkish | 11.04% | 10.09% | 10.75% | 10.27% | 14.98% | 9.64% | 7.30% | 7 of 13 | +70% |
| Vietnamese | 10.47% | 9.31% | 10.07% | 8.19% | 21.52% | 8.96% | 2.75% | 8 of 13 | +281% |
Table 7 breaks multilingual accuracy down by language, with the language specified, for 18 languages. Averaged over the 18 languages, Universal-3.6 Pro is third on both sets: behind Google Gemini 3.5 Live and Soniox Realtime on FLEURS, and behind Speechmatics Enhanced and AWS Transcribe on Common Voice, where its mean is within 3% of the lowest. On FLEURS, Universal-3.6 Pro is lowest in French, Hebrew and Swedish and second in five more languages; Google Gemini 3.5 Live is lowest in 14 of the 18. On Common Voice, Universal-3.6 Pro is lowest in German, French and Korean and second in seven more.
Table 8. Code-switched speech by language pair: normalized WER on the public Common Voice and FLEURS code-switching sets, for Universal-3.6 Pro and the six other systems with the lowest mean error. The last two columns give Universal-3.6 Pro’s rank among all systems and its relative difference from the best other system in the row, including systems not shown.
| metric | Universal-3.6 Pro | AWS Transcribe | Google Gemini 3.5 Live | Reson8 | Soniox Realtime | ElevenLabs Scribe v2 | Meta Muse Voice Transcribe | Rank | vs closest |
|---|---|---|---|---|---|---|---|---|---|
| Mean of five language pairs | 6.46% | 8.30% | 8.84% | 10.08% | 10.77% | 12.37% | 12.85% | 1 of 21 | −22% |
| English–German | 6.13% | 8.12% | 8.29% | 9.73% | 10.81% | 12.29% | 12.31% | 1 of 21 | −24% |
| English–Spanish | 5.91% | 7.17% | 7.83% | 9.61% | 9.67% | 11.06% | 10.91% | 1 of 21 | −18% |
| English–French | 7.23% | 9.33% | 10.05% | 12.60% | 12.15% | 13.16% | 14.64% | 1 of 21 | −23% |
| English–Italian | 6.03% | 7.05% | 8.25% | 8.77% | 10.04% | 11.53% | 11.56% | 1 of 21 | −14% |
| English–Portuguese | 7.00% | 9.81% | 9.79% | 9.67% | 11.18% | 13.79% | 14.85% | 1 of 21 | −28% |
BestWorstCells are coloured by rank within each row (solid blue is best, white is worst). Lower is better in every row.
Speakers who mix English with another language were one of the gaps customers asked us to close (Section 1). Table 8 breaks code-switched speech down by language pair. Universal-3.6 Pro has the lowest error of all systems on every pair, from 14% below AWS Transcribe on English–Italian to 28% below Reson8 on English–Portuguese, and 22% below AWS Transcribe on the mean of the five pairs.
The benchmarks page has the full tables, with every system, entity class, speaker group and language.
7Availability
Universal-3.6 Pro Realtime is available now in the streaming API. Voice focus is off by default; turn it on only when a competing talker is present. Entity-aware endpointing also works with Universal-3.5 Pro Realtime. Our documentation describes how to enable each.