Chariot 发布 Chariot-1-ASR Realtime，支持 10 种印度语言实时语音识别
Chariot Launches Speech Recognition Model for 10 Indian Languages
Chariot 于 10 月 22 日发布实时语音识别模型 Chariot-1-ASR Realtime，支持印地语、马拉地语、孟加拉语、泰米尔语等 10 种印度语言及印度英语，已在 Chariot 平台上线。公司称在 79 个公开测试集、约 15.3 万条录音上测试，批处理模式下错误率比 Google Gemini 3.5 Transcribe 低 26%、比 Sarvam Saaras v4 低 9%；流式模式下比 Sarvam Saaras v4 低 17%、比 ElevenLabs Scribe v2 Realtime 低 57%，这些结果均来自 Chariot 自测。模型提供均衡与高精度两种模式，可自动识别语言，并支持将借用的英文词写成英文字母或对应语言文字。
正文
Chariot, a frontier AI research lab, launched Chariot-1-ASR Realtime on October 22, a speech recognition model that converts spoken words into text while people are still talking. The company said the model outperformed every streaming and batch transcription system it tested across 10 Indian languages.
The model supports Hindi, Marathi, Bengali, Tamil, Telugu, Kannada, Malayalam, Gujarati, Punjabi and Odia. It also supports Indian English and is available on the Chariot platform, according to the company.
Chariot said it tested the model on 79 public test sets from 11 collections, covering around 153,000 audio recordings. The tests included everyday conversations, telephone calls, read-aloud speech and recordings with background noise.
The research lab said its model made 26% fewer errors than Google Gemini 3.5 Transcribe and 9% fewer than Sarvam’s Saaras v4 when compared with systems that process an entire recording before producing text.
In tests against systems that transcribe speech as it arrives, Chariot reported 17% fewer errors than Sarvam Saaras v4, 34% fewer than Google Gemini 3.5 Transcribe Live and 57% fewer than ElevenLabs Scribe v2 Realtime. Chariot was not told which language was being spoken during these tests, while competing systems were given the language settings that worked best for them.
On Monsoon, a public collection of unscripted telephone conversations, Chariot reported an error rate of 5.04% on Hindi calls, compared with 8.27% for OpenAI’s gpt-transcribe. On Indian English calls, its 3% error rate matched results from Microsoft Azure Speech and Zoom Scribe v2 Pro on the Open ASR Leaderboard.
The company also reported a 17.21% error rate on Vaani, a collection of Hindi recordings from speakers across different districts. Sarvam Saaras v4 recorded 17.86% on the same test. Lower error rates indicate more accurate transcription.
These results come from Chariot’s own tests.
Chariot-1-ASR Realtime offers two modes. The balanced mode produces text faster for applications such as voice assistants and live captions. The high-accuracy mode considers more audio before producing text, making it suitable for call transcripts, meeting notes, and dictation.
The model can identify the language automatically or use a language selected by the application. It can also write borrowed English words in either the English alphabet or the script of the language being spoken.
“Speech recognition has always made you choose: fast or accurate. For Indian languages, Chariot-1-ASR Realtime ends that choice,” Chariot co-founder Parth Sarthi said.
The model is based on Chariot-1, the company’s broader speech model, which has been adapted for transcription. Chariot said it plans to use the same underlying model for other speech tasks, including generating speech and identifying languages.