Arth V2
Your agent can only reply once it knows what the customer said.
Arth V2 gets there in 67 ms on 80% of turns.
25%
Lower Latency Than Soniox v5
Beats the Best Model on Latency
12x
Lower Latency Than Soniox v5
~4x
Lower Latency Than Soniox v5
2.6x
Lower Latency Than Soniox v5
Every engine ran on the same 863 calls. We only counted errors that change what the agent does next, like a wrong amount or a missed "nahi".
Soniox v5
Arth V2
Deepgram Nova 3
5% fewer
Sarvam Saaras v4
18% fewer
Cartesia ink-2
24% fewer
Google Gemini 3.5
26% fewer
OpenAI gpt-live
28% fewer
Google Chirp 3
29% fewer
Deepgram Flux
38% fewer
ElevenLabs Scribe v2
60% fewer
Soniox v5 is 0.08 points ahead at 8.77%, which the benchmark treats as a tie, and slower at 89 ms. SquadStack Conversational Streaming ASR Benchmark, 8 kHz, streaming.
Customers pick up at a shop counter, in an auto or with the TV on. Arth holds up best as the line gets noisier.
Clean to noisy calls
How much each engine's error rate rises with noise
Rise in semantic WER, percentage points, clean to noisy calls.
10.1%
Lowest error rate on calls with heavy Hindi-English mixing, ahead of Soniox (10.5%) and Sarvam (13.6%).
A wrong amount or a missed "nahi" can lose the sale. Arth V2 makes the fewest of these errors against Sarvam, Deepgram, Google and OpenAI.
A wrong amount, date or OTP
A "nahi" heard as "haan"
A promise to pay or buy, misheard
Any other word that changes meaning
Errors as a share of all words spoken, across Arth V2, Deepgram Nova 3, Sarvam Saaras v4, Google Gemini 3.5 and OpenAI gpt-live.
Arth V2 is trained on real Indian sales calls, with the accents, noise, language switches, amounts and OTPs that decide a sale.
60 crore+
60 crore+
10 crore+
Unique Speakers
85%
of India's Pincodes
1 crore+
Live Minutes in Just 2 Months
863 real Hindi-English sales calls from 19 campaigns across five industries. Every error, for every engine, is published call by call.
Calls in Every Mix of Noise and Hinglish
All nine combinations are covered, and no industry holds more than 23% of the speech, so no single one decides the ranking.
Semantic WER
Share of words an engine gets wrong in a way that changes what the agent does. A wrong EMI amount counts. An extra "haan" does not.
FTR at p80
Time from the end of the customer's turn to the final transcript, on 80% of turns.
Streaming, on 8 kHz Audio
Every engine heard every call live, on normal phone audio.