Overall experience
I enjoyed talking to this avatar and found the overall experience positive.
Avatar quality
benchmark
How natural does it feel to talk to an avatar? Participants rated six aspects of the experience, from lip sync and appearance to response timing.
72.9%of Anam’s ratings were positiveAcross all six categories · full pipelines
6categories rated after each conversationFrom visual quality to overall experience
- Measured
- Data from 60 participants
- Includes 960 conversations · 5,760 ratings
- Anam · Cara-4.1 Express
- Tavus · phoenix-4.5
- HeyGen
- LemonSlice
Providers’ full pipelines
Human ratings: How people scored each product.
Each provider ran its complete conversation pipeline, with Anam using Cara-4.1 Express and competitors running their defaults. Anam received positive scores on 72.9% of ratings, compared with 54.1% for Tavus, 52.1% for HeyGen and 51.5% for LemonSlice.
Positive ratings across all six categories
Providers’ full pipelines · Higher is better
- Anam72.9%
- Tavus54.1%
- HeyGen52.1%
- LemonSlice51.5%
The five-point scale: −2 Strongly disagree · −1 Disagree · 0 Neutral · +1 Agree · +2 Strongly agree.
| Vendor | -2 | -1 | 0 | +1 | +2 | Positive |
|---|---|---|---|---|---|---|
| AnamCara-4.1 Express | 2% | 10% | 15% | 51% | 22% | 72.9% |
| Tavusphoenix-4.5 | 7% | 18% | 21% | 41% | 13% | 54.1% |
| HeyGenDefault | 6% | 21% | 21% | 38% | 14% | 52.1% |
| LemonSliceDefault | 10% | 19% | 20% | 38% | 13% | 51.5% |
The five score columns are rounded to whole percentages. The positive total is calculated from the underlying ratings and shown to one decimal place, so it can differ from adding the rounded +1 and +2 columns.
Providers’ full pipelines
Six categories: From how an avatar looks to how a conversation feels.
Each cell shows the share of positive answers, followed by the mean score on the −2 to +2 scale. Every category has 240 answers per provider. “Overall experience” is its own question; it is not an average of the other five.
| Category | Anam | Tavus | HeyGen | LemonSlice |
|---|---|---|---|---|
| Overall experience | 72%Mean +0.73 | 54%Mean +0.33 | 47%Mean +0.25 | 49%Mean +0.21 |
| Lip sync quality | 68%Mean +0.72 | 40%Mean +0.09 | 38%Mean -0.10 | 30%Mean -0.32 |
| Visual quality | 69%Mean +0.75 | 66%Mean +0.64 | 51%Mean +0.37 | 53%Mean +0.33 |
| Naturalness | 65%Mean +0.61 | 45%Mean +0.15 | 32%Mean -0.23 | 38%Mean -0.02 |
| Response time (participant rating) | 90%Mean +1.16 | 60%Mean +0.41 | 78%Mean +0.88 | 73%Mean +0.71 |
| Interruptibility | 74%Mean +0.90 | 59%Mean +0.50 | 68%Mean +0.80 | 65%Mean +0.64 |
What participants rated
Lip sync quality
The mouth and facial movements matched the spoken audio.
Visual quality
The appearance of the avatar was visually pleasing and free of obvious flaws.
Naturalness
The avatar's behavior seemed natural and lifelike.
Response time (participant rating)
The avatar’s response time felt natural and as expected in conversation.
Interruptibility
I could interrupt the avatar as I would in a natural conversation.
Anam had the highest mean score in all six categories. Paired comparisons with the highest-scoring competitor in each category found a statistically significant lead for overall experience, lip sync, naturalness and responsiveness (p < 0.001).
The visual-quality comparison with Tavus (p = 0.182) and interruptibility comparison with HeyGen (p = 0.261) did not establish a lead. A higher percentage alone does not show that a difference is statistically significant.
Responsiveness here describes how the timing felt to raters, not measured reply or connection times.
Providers’ full pipelines
Participant scores: 53 of 60 scored Anam above the next-highest-rated provider.
We averaged each participant’s answers across all six categories and their conversations with each provider, then compared those averages within the same participant. This is preference inferred from their scores, rather than a separate forced-choice question.
| Comparison | Scored Anam higher | Scored competitor higher | Scored equally |
|---|---|---|---|
| Anam vs Tavus | 53 | 6 | 1 |
| Anam vs HeyGen | 51 | 7 | 2 |
| Anam vs LemonSlice | 55 | 4 | 1 |
These counts and the positive-rating percentages summarize the same study at different levels: one counts participants, the other counts individual answers. They are not independent experiments.
How do the avatar models compare on the same pipeline?
A separate study used 60 participants and 1,200 conversations, with 20 conversations per participant. Every provider used the same speech-to-text model, language model and text-to-speech model. The avatar provider changed.
The shared LiveKit pipeline used Deepgram Nova-3 for speech recognition, Groq-hosted GPT OSS 120B as the language model, and Cartesia Sonic-3.6 for speech generation.
Synthesia is included in this study because it uses a customer-supplied speech-and-language pipeline. This provides a like-for-like comparison of that setup. The two studies are reported separately because their configurations differ.
Providers’ avatar models on LiveKit
Human ratings: Anam led with the pipeline held constant.
Anam received positive scores on 66.8% of ratings, compared with 54.6% for Synthesia, the next highest provider. This study collected 7,200 answers across all six categories on October 16–17, 2026.
Positive ratings across all six categories
Shared LiveKit pipeline · Higher is better
- Anam66.8%
- Synthesia54.6%
- Tavus54.0%
- HeyGen43.8%
- LemonSlice41.2%
| Vendor | -2 | -1 | 0 | +1 | +2 | Positive |
|---|---|---|---|---|---|---|
| AnamCara-4.1 Express | 4% | 11% | 18% | 44% | 23% | 66.8% |
| SynthesiaDefault | 6% | 18% | 22% | 40% | 15% | 54.6% |
| Tavusphoenix-4.5 | 8% | 18% | 20% | 38% | 16% | 54.0% |
| HeyGenDefault | 8% | 24% | 24% | 31% | 12% | 43.8% |
| LemonSliceDefault | 12% | 25% | 22% | 29% | 12% | 41.2% |
As above, the five score columns are rounded independently; positive totals use the underlying ratings.
| Category | Anam | Synthesia | Tavus | HeyGen | LemonSlice |
|---|---|---|---|---|---|
| Overall experience | 70%Mean +0.71 | 56%Mean +0.44 | 50%Mean +0.35 | 39%Mean +0.17 | 35%Mean 0.00 |
| Lip sync quality | 65%Mean +0.63 | 48%Mean +0.26 | 36%Mean -0.06 | 31%Mean -0.25 | 32%Mean -0.19 |
| Visual quality | 71%Mean +0.82 | 49%Mean +0.25 | 74%Mean +0.80 | 56%Mean +0.46 | 60%Mean +0.42 |
| Naturalness | 62%Mean +0.54 | 54%Mean +0.33 | 48%Mean +0.16 | 22%Mean -0.42 | 37%Mean -0.16 |
| Response time (participant rating) | 73%Mean +0.80 | 63%Mean +0.56 | 57%Mean +0.38 | 59%Mean +0.46 | 39%Mean -0.09 |
| Interruptibility | 60%Mean +0.66 | 57%Mean +0.55 | 59%Mean +0.55 | 56%Mean +0.57 | 43%Mean +0.25 |
Tavus had a higher share of positive visual-quality ratings (74% versus 71%), while Anam’s mean was slightly higher (+0.82 versus +0.80). The mean includes how strongly people agreed or disagreed. The paired comparison did not establish a visual-quality lead over Tavus (p = 0.824).
Anam’s mean-score lead over the best competitor was significant for overall experience, lip sync, naturalness and perceived responsiveness (p < 0.05). Interruptibility did not establish a lead over HeyGen (p = 0.166). These ratings describe how response timing felt, not measured response latency.
Providers’ avatar models on LiveKit
Participant scores: 42 of 60 scored Anam above the next-highest-rated provider.
Using the same comparison of each participant’s average scores, 42 scored Anam higher than the next-highest-rated provider, 15 scored that provider higher and 3 scored them equally.
| Comparison | Scored Anam higher | Scored competitor higher | Scored equally |
|---|---|---|---|
| Anam vs Synthesia | 42 | 15 | 3 |
| Anam vs Tavus | 46 | 10 | 4 |
| Anam vs HeyGen | 48 | 8 | 4 |
| Anam vs LemonSlice | 54 | 3 | 3 |
Methodology: Live conversations, rated by people.
Participants played 20 Questions with the avatars in a browser, using their microphone. They chose a word and the avatar tried to guess it through conversation, with a two-minute session cap. After each conversation, participants answered the same six questions.
A shared task
Each participant tried every provider in their study. Four avatar slots per provider covered clothing, food, professions and sports. Practice rounds were excluded from the results.
A consistent scale
Answers ranged from Strongly disagree (−2) to Strongly agree (+2). Positive shares count +1 and +2 answers. Mean scores retain all five possible answers, including neutral and negative scores.
Paired comparisons
Statistical comparisons use each participant’s own average difference between providers. The participant is the unit of observation; repeated answers are not treated as independent people. The study uses p < 0.05 as its significance threshold.
These results describe the tested avatars, task and configurations in October 2026. A score reflects the full experience the participant saw and heard. It does not isolate resolution, voice, appearance or any other single feature. Products and performance can change.
Data and transparency
The public benchmark repository includes the evaluation code, methodology, de-identified ratings and provider configurations. You can reproduce the published results offline with Python, without API keys or provider accounts.
Anam conducted this benchmark, and we recognise the conflict of interest in comparing our own product with competitors. The results include every tested provider in each setup.
Build conversations that feel natural
Get started with Anam’s real-time avatar API.