Datasets:
Leaderboard
Written from the scored output of the benchmark run. That output is the single source of truth; this file and the figures derive from it. Pooled over the 863 calls and about 67,000 reference words (67,388 to 67,393 per system).
Ordering. Every table is ordered ascending by its own metric (lowest, best, first). ★ marks SquadStack's in-house systems: they are ranked with the others, not separated, and marked because a model tuned on this kind of audio is not a like-for-like peer of a general-purpose endpoint.
Reading the numbers. These are point estimates. The top four systems are within about 0.5 points of each other and should not be ranked on this table alone. Confidence intervals and paired comparisons are planned for the next release.
Overall
| System | Vendor | Category | Semantic WER | Raw WER | Share of raw errors that change meaning |
|---|---|---|---|---|---|
| Soniox v5 | Soniox | Commercial | 8.77 | 27.29 | 32% |
| Arth V2 ★ | SquadStack | In-house | 8.85 | 29.14 | 30% |
| Arth V1 ★ | SquadStack | In-house | 9.17 | 29.83 | 31% |
| Nova 3 | Deepgram | Commercial | 9.31 | 30.98 | 30% |
| Saaras v4 | Sarvam | Commercial | 10.84 | 33.25 | 33% |
| Cartesia ink-2 | Cartesia | Commercial | 11.58 | 36.27 | 32% |
| Gemini 3.5 Transcribe | Commercial | 11.93 | 35.82 | 33% | |
| OpenAI gpt-live-transcribe | OpenAI | Commercial | 12.30 | 38.14 | 32% |
| Chirp 3 | Commercial | 12.47 | 34.03 | 37% | |
| Flux | Deepgram | Commercial | 14.33 | 37.76 | 38% |
| Scribe v2 Realtime | ElevenLabs | Commercial | 22.09 | 54.95 | 40% |
Raw WER by edit type
The three components sum to raw WER (percent of reference words). A deletion-heavy system drops speech; an insertion-heavy one invents it.
| System | Raw WER | Substitutions | Deletions | Insertions |
|---|---|---|---|---|
| Soniox v5 | 27.29 | 12.66 | 5.45 | 9.17 |
| Arth V2 ★ | 29.14 | 13.17 | 8.05 | 7.92 |
| Arth V1 ★ | 29.83 | 13.42 | 8.96 | 7.45 |
| Nova 3 | 30.98 | 11.69 | 14.49 | 4.80 |
| Saaras v4 | 33.25 | 14.73 | 6.26 | 12.26 |
| Chirp 3 | 34.03 | 15.14 | 10.96 | 7.93 |
| Gemini 3.5 Transcribe | 35.82 | 14.48 | 14.67 | 6.68 |
| Cartesia ink-2 | 36.27 | 14.69 | 3.84 | 17.73 |
| Flux | 37.76 | 18.28 | 12.91 | 6.56 |
| OpenAI gpt-live-transcribe | 38.14 | 18.16 | 13.52 | 6.46 |
| Scribe v2 Realtime | 54.95 | 23.99 | 9.09 | 21.87 |
Rank: raw WER against semantic WER
| System | Raw rank | Semantic rank | Change |
|---|---|---|---|
| Soniox v5 | 1 | 1 | 0 |
| Arth V2 ★ | 2 | 2 | 0 |
| Arth V1 ★ | 3 | 3 | 0 |
| Nova 3 | 4 | 4 | 0 |
| Saaras v4 | 5 | 5 | 0 |
| Cartesia ink-2 | 8 | 6 | +2 |
| Gemini 3.5 Transcribe | 7 | 7 | 0 |
| OpenAI gpt-live-transcribe | 10 | 8 | +2 |
| Chirp 3 | 6 | 9 | -3 |
| Flux | 9 | 10 | -1 |
| Scribe v2 Realtime | 11 | 11 | 0 |
Semantic WER by code-mixing (from the human reference)
| System | low_mix n=139 |
mid_mix n=460 |
high_mix n=264 |
|---|---|---|---|
| Soniox v5 | 8.6 | 8.2 | 10.4 |
| Arth V2 ★ | 8.0 | 8.7 | 9.7 |
| Arth V1 ★ | 8.0 | 8.8 | 10.8 |
| Nova 3 | 8.0 | 9.1 | 10.8 |
| Saaras v4 | 11.0 | 9.8 | 13.6 |
| Cartesia ink-2 | 12.1 | 10.4 | 14.5 |
| Gemini 3.5 Transcribe | 11.1 | 11.1 | 14.8 |
| OpenAI gpt-live-transcribe | 11.5 | 11.7 | 14.4 |
| Chirp 3 | 10.9 | 11.9 | 15.2 |
| Flux | 13.0 | 13.4 | 17.8 |
| Scribe v2 Realtime | 21.9 | 21.1 | 24.9 |
Semantic WER by acoustic condition
| System | clean n=107 |
moderate n=509 |
noisy n=247 |
|---|---|---|---|
| Soniox v5 | 4.9 | 9.4 | 12.7 |
| Arth V2 ★ | 6.7 | 9.4 | 10.5 |
| Arth V1 ★ | 7.0 | 9.5 | 11.5 |
| Nova 3 | 6.2 | 10.0 | 12.0 |
| Saaras v4 | 6.7 | 11.6 | 14.7 |
| Cartesia ink-2 | 5.6 | 12.8 | 17.1 |
| Gemini 3.5 Transcribe | 7.8 | 12.8 | 15.7 |
| OpenAI gpt-live-transcribe | 8.8 | 13.0 | 15.4 |
| Chirp 3 | 7.3 | 14.0 | 16.0 |
| Flux | 11.2 | 15.4 | 16.2 |
| Scribe v2 Realtime | 17.2 | 23.1 | 26.5 |
The code-mixing and noise margins overlap (the grid is unbalanced). Read the nine-cell grid in the dataset card for the joint view.
Semantic WER by domain
| System | BFSI n=122 |
Education n=54 |
Logistics n=260 |
Marketplace n=240 |
Travel n=187 |
|---|---|---|---|---|---|
| Soniox v5 | 8.0 | 6.5 | 8.1 | 13.9 | 7.9 |
| Arth V2 ★ | 8.3 | 7.3 | 8.2 | 12.2 | 8.6 |
| Arth V1 ★ | 9.1 | 6.9 | 8.2 | 12.8 | 8.7 |
| Nova 3 | 9.3 | 6.1 | 8.7 | 13.6 | 8.5 |
| Saaras v4 | 12.7 | 6.8 | 10.2 | 14.2 | 8.5 |
| Cartesia ink-2 | 11.8 | 6.2 | 12.9 | 16.2 | 9.8 |
| Gemini 3.5 Transcribe | 13.2 | 7.2 | 12.1 | 15.4 | 10.0 |
| OpenAI gpt-live-transcribe | 13.2 | 8.4 | 11.3 | 17.5 | 10.1 |
| Chirp 3 | 13.9 | 7.9 | 13.1 | 16.2 | 9.6 |
| Flux | 14.4 | 12.0 | 13.1 | 19.3 | 12.8 |
| Scribe v2 Realtime | 24.3 | 16.0 | 19.5 | 26.0 | 21.9 |
Education is 54 calls; read it as a direction, not a number.
Semantic WER by speaker gender
| System | Female n=146 |
Male n=717 |
|---|---|---|
| Soniox v5 | 8.6 | 8.8 |
| Arth V2 ★ | 8.3 | 9.0 |
| Arth V1 ★ | 8.3 | 9.3 |
| Nova 3 | 8.1 | 9.5 |
| Saaras v4 | 10.4 | 10.9 |
| Cartesia ink-2 | 10.4 | 11.8 |
| Gemini 3.5 Transcribe | 11.0 | 12.1 |
| OpenAI gpt-live-transcribe | 11.6 | 12.4 |
| Chirp 3 | 11.9 | 12.6 |
| Flux | 12.7 | 14.7 |
| Scribe v2 Realtime | 22.4 | 22.0 |
There are 146 female calls (16.9%), 12 of them labelled from the model vote. The gender comparison is indicative only.
Semantic WER by error class
The judge assigns each meaning-changing error a class. The classes partition the semantic errors, so each row sums to its semantic WER (percent of reference words). A wrong number or a flipped negation changes the outcome of a call; a content error usually costs a retry.
| System | Semantic WER | Number | Negation | Name | Commitment | Content | Other |
|---|---|---|---|---|---|---|---|
| Soniox v5 | 8.77 | 1.07 | 0.43 | 0.77 | 0.30 | 6.18 | 0.00 |
| Arth V2 ★ | 8.85 | 1.10 | 0.52 | 0.92 | 0.42 | 5.89 | 0.00 |
| Arth V1 ★ | 9.17 | 1.11 | 0.52 | 0.99 | 0.41 | 6.14 | 0.01 |
| Nova 3 | 9.31 | 1.24 | 0.55 | 1.11 | 0.48 | 5.92 | 0.00 |
| Saaras v4 | 10.84 | 1.29 | 0.59 | 0.90 | 0.50 | 7.55 | 0.01 |
| Cartesia ink-2 | 11.58 | 1.15 | 0.60 | 0.93 | 0.42 | 8.47 | 0.00 |
| Gemini 3.5 Transcribe | 11.93 | 1.43 | 0.79 | 0.95 | 0.65 | 8.08 | 0.01 |
| OpenAI gpt-live-transcribe | 12.30 | 1.59 | 0.68 | 1.02 | 0.46 | 8.55 | 0.00 |
| Chirp 3 | 12.47 | 1.91 | 0.70 | 0.91 | 0.52 | 8.41 | 0.02 |
| Flux | 14.33 | 1.68 | 0.86 | 1.36 | 0.64 | 9.79 | 0.01 |
| Scribe v2 Realtime | 22.09 | 2.11 | 1.33 | 1.47 | 0.92 | 16.25 | 0.02 |
Judge validation
Semantic WER is judged by an LLM, so its agreement with human adjudication is part of the result. This table is to be measured on a stratified sample of errors with two annotators and is not part of version 1.0.0:
| Sample size | to be measured |
| Agreement between judge and human adjudication | to be measured |
| Per-call correlation | to be measured |
Latency
FTR is the time from our end-of-turn finalize signal (sent at each production voice-activity stop) to the vendor's final reply for that turn. The table gives the 80th percentile over all turns, ordered by it. Measured on a locked 50-call set (70 minutes of audio, stratified by duration), streamed in real time at 5 concurrency from one server in Delhi, one vendor at a time. Semantic WER is from the 863-call run.
Full method, P50–P99 and TTFS per system: LATENCY.md.
| System | Region | Hosting | FTR P80 (ms) | Semantic WER |
|---|---|---|---|---|
| Arth V2 ★ | India | Self-hosted | 67.0 | 8.85 |
| Arth V1 ★ | India | Self-hosted | 67.0 | 9.17 |
| Scribe v2 Realtime | India | Provider cloud | 86.5 | 22.09 |
| Soniox v5 | India | Provider cloud | 89.4 | 8.77 |
| Nova 3 | India | Provider cloud | 93.8 | 9.31 |
| Flux | India | Provider cloud | 124.4 | 14.33 |
| Saaras v4 | India | Provider cloud | 176.7 | 10.84 |
| Cartesia ink-2 | India | Provider cloud | 183.6 | 11.58 |
| Gemini 3.5 Transcribe | not documented | Provider cloud | 264.2 | 11.93 |
| OpenAI gpt-live-transcribe | not documented | Provider cloud | 830.8 | 12.30 |
Notes:
- Soniox v5 is measured on its India regional endpoint; the Japan, EU and US endpoints are compared in LATENCY.md (Japan gave an FTR P80 of 236.2 ms).
- Sarvam and Gemini do not tag their reply, so their FTR is inferred as the first final received after the finalize signal.
- OpenAI streams the turn text about 400 ms before the final event that FTR measures.
- Chirp 3 does not support a client-side finalize, so FTR does not apply.
Systems and versions
All 11 systems were scored on the same 863 recordings through the same pipeline. Model identifiers: Arth V1 and V2 (SquadStack), Deepgram Nova 3 and Flux, Soniox stt-rt-v5, Sarvam Saaras v4, ElevenLabs Scribe v2 Realtime, Cartesia ink-2, Google Chirp 3, Gemini 3.5 Transcribe (live), OpenAI gpt-live-transcribe.
Submitting
See SUBMISSION.md.