SquadStack.ai 发布 Arth V2 语音转文字模型，宣称 67ms 转写延迟
SquadStack.ai Launches Arth V2 STT Model, Claims 67ms Transcription Latency
SquadStack.ai 发布流式语音转文字模型 Arth V2，在 863 通真实印度销售电话基准上语义词错误率为 8.85%，80% 情况下在客户说完后 67 毫秒内返回最终转写。其含义改变类错误比 Sarvam Saaras v4 少 18%、比 Google Gemini 3.5 Transcribe 少 26%、比 OpenAI gpt-live-transcribe 少 28%。该模型已用于其企业语音 AI 智能体，并以封闭测试形式向其他团队开放。
正文
当前显示原文，中文全文译文尚未就绪。
SquadStack.ai has launched Arth V2, claiming its streaming speech-to-text model achieved the lowest latency among engines tested on the company's benchmark of Indian telephone conversations. The company reported a semantic word error rate of 8.85% on a dataset of 863 real sales calls.
Arth V2 is designed to transcribe Indian sales conversations over telephone networks, including understanding code-switching. SquadStack.ai said the model processed more than 1 crore minutes of live customer calls in two months and handles 80% of the more than 50 lakh AI calls it makes daily.
“Speech-to-text looks solved, but it is still the biggest point of failure for a voice AI agent. Feed a language model the wrong words and it gives the wrong answer,” Apurv Agrawal, Co-founder and CEO of SquadStack.ai, said in a statement.
According to SquadStack.ai, Arth V2 recorded a semantic word error rate of 8.85%, compared with 8.77% for Soniox v5. The company said Arth V2 returned its final transcript within 67 milliseconds of a customer finishing a speaking turn in 80% of cases. It claimed no other engine tested matched its combination of accuracy and speed.
SquadStack.ai also reported that Arth V2 made 18% fewer meaning-changing errors than Sarvam Saaras v4, 26% fewer than Google Gemini 3.5 Transcribe, and 28% fewer than OpenAI gpt-live-transcribe.
The benchmark uses 8-kilohertz narrowband telephone audio from 863 unscripted Indian sales calls. SquadStack.ai said more than four in five minutes of speech are code-mixed and nearly nine in 10 calls contain background noise. The scoring method focuses on transcription errors that change the intended meaning, such as incorrectly recognising a loan amount or payment date.
According to SquadStack.ai, the model was trained in-house from scratch on more than 60 crore minutes of Indian sales conversations involving over 10 crore unique speakers.
Vikas Gulati, Co-founder and CTO of SquadStack.ai, said the company plans to update Arth as its performance improves. “That is why we run SquadStack.ai as a model factory, and ship a new version whenever it beats what is in production,” he said.
SquadStack.ai said Arth V2 runs within its voice AI agents for enterprise customers and is available in closed beta for other teams that want to test it on their own calls.
The company has also made the benchmark's test audio, reference transcripts, and scoring method available on Hugging Face.