Speech-to-text (STT)

Model details on this page come from each provider’s official documentation and announcements, checked on October 16, 2026. Providers update models, benchmarks, language support and availability often. Always check the linked official source for the latest information before choosing a model.

Speech-to-text models transcribe the caller’s speech so the LLM can respond to it. Select one in the Transcript Model field of the Voice Agent configuration.

At a glance

Model	Provider	Model ID	Highlights	Official docs
Nova 3	Deepgram	`nova-3-general`	General-purpose STT, real-time code-switching across 10 languages	Deepgram
Flux (English)	Deepgram	`flux-general-en`	Built for voice agents, with built-in end-of-turn detection	Deepgram
Flux (Multilingual)	Deepgram	`flux-general-multi`	Flux for 10 languages, with optional language hints	Deepgram
GPT Live Transcribe	OpenAI	`gpt-live-transcribe`	Low-latency streaming transcription with keyword and language hints	OpenAI
GPT Realtime Whisper	OpenAI	`gpt-realtime-whisper`	Streaming transcription (legacy)	OpenAI
Telephony	Google	`telephony`	Tuned for 8 kHz phone-call audio	Google Cloud
Chirp 3	Google	`chirp_3`	Multilingual, with diarization and automatic language detection	Google Cloud
Scribe v2 Realtime	ElevenLabs	`scribe_v2_realtime`	~150 ms live transcription, 90+ languages	ElevenLabs
Saaras v3	Sarvam	`saaras:v3`	Indian languages, 22 Indian languages plus English	Sarvam
Saarika v2.5	Sarvam	`saarika:v2.5`	Indian languages (being deprecated)	Sarvam

Deepgram

Nova 3

nova-3-general · Official docs Deepgram describes Nova 3 as its “highest-performing general-purpose ASR”. It is recommended for meetings, multi-speaker, multilingual, noisy or far-field audio, in batch or streaming. Nova 3 does not include turn detection.

Metric	Value	Source
Median word error rate (streaming)	6.84%, vs 14.92% for the next-best competitor	Deepgram
Median word error rate (batch)	5.26%	Deepgram
Test set	2,703 audio files (81.69 hours) across multiple domains, including phone calls, drive-thru, conversational AI and voicemail	Deepgram
Code-switching	Real time across 10 languages (English, Spanish, French, German, Hindi, Russian, Portuguese, Japanese, Italian, Dutch)	Deepgram
Keyterm prompting	Up to 100 key terms	Deepgram

Languages available in SubVerse AI (47)

Multilingual (auto-detect) (multi)
Arabic (ar)
Arabic (Kuwait) (ar-KW)
Arabic (Saudi Arabia) (ar-SA)
Arabic (UAE) (ar-AE)
Bengali (bn-IN)
Bulgarian (bg-BG)
Croatian (hr-HR)
Czech (cs-CZ)
Danish (da-DK)
Dutch (nl-NL)
English (Australia) (en-AU)
English (Canada) (en-CA)
English (India) (en-IN)
English (UK) (en-GB)
English (US) (en-US)
Filipino (fil-PH)
Finnish (fi-FI)
French (Canada) (fr-CA)
French (France) (fr-FR)
German (de-DE)
German (Switzerland) (de-CH)
Greek (el-GR)
Hindi (hi-IN)
Hungarian (hu-HU)
Indonesian (id-ID)
Italian (it-IT)
Japanese (ja-JP)
Kannada (kn-IN)
Korean (ko-KR)
Malay (ms-MY)
Marathi (mr-IN)
Norwegian (no-NO)
Polish (pl-PL)
Portuguese (Brazil) (pt-BR)
Portuguese (Portugal) (pt-PT)
Romanian (ro-RO)
Russian (ru-RU)
Slovak (sk-SK)
Spanish (Mexico) (es-MX)
Spanish (Spain) (es-ES)
Swedish (sv-SE)
Tamil (ta-IN)
Telugu (te-IN)
Turkish (tr-TR)
Ukrainian (uk-UA)
Vietnamese (vi-VN)

Flux (English)

flux-general-en · Official docs Deepgram calls Flux “the first conversational speech recognition model built specifically for voice agents”. It has end-of-turn detection built into the model, configurable turn-taking, and “Nova-3 level accuracy”.

Metric	Value	Source
End-of-turn detection latency	~260 ms	Deepgram
Agent response latency	200–600 ms lower than pipeline approaches	Deepgram
False interruptions	~30% fewer	Deepgram
End-of-turn timing	Fires within 1.5 s, 95% of the time	Deepgram
Accuracy	Word error rate comparable to Nova 3 Streaming	Deepgram

Languages available in SubVerse AI (7)

English (Australia) (en-AU)
English (Canada) (en-CA)
English (India) (en-IN)
English (Philippines) (en-PH)
English (Singapore) (en-SG)
English (UK) (en-GB)
English (US) (en-US)

Flux (Multilingual)

flux-general-multi · Official docs Deepgram says this model “extends Flux to 10 languages with optional language_hint biasing”: English, Spanish, French, German, Hindi, Russian, Portuguese, Japanese, Italian and Dutch. It has the same end-of-turn detection and barge-in handling as Flux (English). Deepgram’s Flux launch benchmarks (see Flux (English)) were published before the multilingual variant. No separate figures were found for this variant from Deepgram or from an independent source.

Languages available in SubVerse AI (19)

Dutch (nl-NL)
English (Australia) (en-AU)
English (Canada) (en-CA)
English (India) (en-IN)
English (Philippines) (en-PH)
English (Singapore) (en-SG)
English (UK) (en-GB)
English (US) (en-US)
French (Canada) (fr-CA)
French (France) (fr-FR)
German (de-DE)
Hindi (hi-IN)
Italian (it-IT)
Japanese (ja-JP)
Portuguese (Brazil) (pt-BR)
Portuguese (Portugal) (pt-PT)
Russian (ru-RU)
Spanish (Mexico) (es-MX)
Spanish (Spain) (es-ES)

OpenAI

GPT Live Transcribe

gpt-live-transcribe · Official docs OpenAI describes it as “a streaming speech-to-text model for applications that need low-latency transcript deltas from live audio”. It supports tunable latency, free-form context, keyword hints and multiple language hints. OpenAI’s realtime transcription guide recommends starting with this model.

Detail	Value	Source
Latency setting	`delay` of `minimal`, `low`, `medium`, `high` or `xhigh`. Higher delay “can improve word error rate”.	OpenAI
Not supported	Word-level timestamps, speaker labels, confidence scores, detected-language output	OpenAI
Benchmarks	Not published. No independent figures found.	—

Languages available in SubVerse AI (44)

Arabic (Kuwait) (ar-KW)
Arabic (Saudi Arabia) (ar-SA)
Arabic (UAE) (ar-AE)
Bulgarian (bg-BG)
Chinese, Mandarin (Traditional, Taiwan) (cmn-Hant-TW)
Croatian (hr-HR)
Czech (cs-CZ)
Danish (da-DK)
Dutch (nl-NL)
English (Australia) (en-AU)
English (Canada) (en-CA)
English (India) (en-IN)
English (UK) (en-GB)
English (US) (en-US)
Filipino (fil-PH)
Finnish (fi-FI)
French (Canada) (fr-CA)
French (France) (fr-FR)
German (de-DE)
Greek (el-GR)
Hindi (hi-IN)
Hungarian (hu-HU)
Indonesian (id-ID)
Italian (it-IT)
Japanese (ja-JP)
Kannada (kn-IN)
Korean (ko-KR)
Malay (ms-MY)
Marathi (mr-IN)
Norwegian (no-NO)
Polish (pl-PL)
Portuguese (Brazil) (pt-BR)
Portuguese (Portugal) (pt-PT)
Romanian (ro-RO)
Russian (ru-RU)
Slovak (sk-SK)
Spanish (Mexico) (es-MX)
Spanish (Spain) (es-ES)
Spanish (US) (es-US)
Swedish (sv-SE)
Tamil (ta-IN)
Turkish (tr-TR)
Ukrainian (uk-UA)
Vietnamese (vi-VN)

GPT Realtime Whisper

gpt-realtime-whisper · Official docs A streaming speech-to-text model for low-latency transcripts from live audio.

OpenAI’s migration guide treats GPT Realtime Whisper as a legacy model and recommends migrating live transcription to GPT Live Transcribe. See OpenAI’s migration guide.

Detail	Value	Source
Context window	16,000 tokens	OpenAI
Max output	2,000 tokens	OpenAI
Benchmarks	Not published. No independent figures found.	—

Languages available in SubVerse AI (44)

Arabic (Kuwait) (ar-KW)
Arabic (Saudi Arabia) (ar-SA)
Arabic (UAE) (ar-AE)
Bulgarian (bg-BG)
Chinese, Mandarin (Traditional, Taiwan) (cmn-Hant-TW)
Croatian (hr-HR)
Czech (cs-CZ)
Danish (da-DK)
Dutch (nl-NL)
English (Australia) (en-AU)
English (Canada) (en-CA)
English (India) (en-IN)
English (UK) (en-GB)
English (US) (en-US)
Filipino (fil-PH)
Finnish (fi-FI)
French (Canada) (fr-CA)
French (France) (fr-FR)
German (de-DE)
Greek (el-GR)
Hindi (hi-IN)
Hungarian (hu-HU)
Indonesian (id-ID)
Italian (it-IT)
Japanese (ja-JP)
Kannada (kn-IN)
Korean (ko-KR)
Malay (ms-MY)
Marathi (mr-IN)
Norwegian (no-NO)
Polish (pl-PL)
Portuguese (Brazil) (pt-BR)
Portuguese (Portugal) (pt-PT)
Romanian (ro-RO)
Russian (ru-RU)
Slovak (sk-SK)
Spanish (Mexico) (es-MX)
Spanish (Spain) (es-ES)
Spanish (US) (es-US)
Swedish (sv-SE)
Tamil (ta-IN)
Turkish (tr-TR)
Ukrainian (uk-UA)
Vietnamese (vi-VN)

Google

Telephony

telephony · Official docs Google Cloud’s telephony model is built for “audio that originates from an audio phone call, typically recorded at an 8 kHz sampling rate”. Google describes it as ideal for customer service, teleconferencing and automated kiosk applications.

Detail	Value	Source
Intended audio	Phone calls, typically 8 kHz	Google Cloud
Benchmarks	Not published. No independent figures found.	—

Languages available in SubVerse AI (17)

Dutch (nl-NL)
English (Australia) (en-AU)
English (India) (en-IN)
English (UK) (en-GB)
English (US) (en-US)
French (Canada) (fr-CA)
French (France) (fr-FR)
German (de-DE)
Hindi (hi-IN)
Italian (it-IT)
Japanese (ja-JP)
Korean (ko-KR)
Portuguese (Brazil) (pt-BR)
Portuguese (Portugal) (pt-PT)
Spanish (Mexico) (es-MX)
Spanish (Spain) (es-ES)
Spanish (US) (es-US)

Chirp 3

chirp_3 · Official docs Chirp 3 is the latest generation of Google’s multilingual speech recognition models. Google says it improves accuracy and speed over earlier Chirp models, and adds speaker diarization and automatic language detection.

Metric	Value	Source
Recognition modes	Streaming, synchronous and batch	Google Cloud
Speech adaptation	Up to 1,000 phrases	Google Cloud
Not supported	Word-level timestamps and word-level confidence	Google Cloud
AA-WER (non-streaming) · Third-party	4.3%	Artificial Analysis

Google does not publish word error rates for Chirp 3. The Third-party figure is published by Artificial Analysis, an independent benchmarking company, not by Google. Accessed October 16, 2026.

Languages available in SubVerse AI (49)

Arabic (Kuwait) (ar-KW)
Arabic (Saudi Arabia) (ar-SA)
Arabic (UAE) (ar-AE)
Bengali (bn-IN)
Bulgarian (bg-BG)
Chinese, Mandarin (Traditional, Taiwan) (cmn-Hant-TW)
Croatian (hr-HR)
Czech (cs-CZ)
Danish (da-DK)
Dutch (nl-NL)
English (Australia) (en-AU)
English (India) (en-IN)
English (UK) (en-GB)
English (US) (en-US)
Filipino (fil-PH)
Finnish (fi-FI)
French (Canada) (fr-CA)
French (France) (fr-FR)
German (de-DE)
Greek (el-GR)
Gujarati (gu-IN)
Hindi (hi-IN)
Hungarian (hu-HU)
Indonesian (id-ID)
Italian (it-IT)
Japanese (ja-JP)
Kannada (kn-IN)
Korean (ko-KR)
Malay (ms-MY)
Malayalam (ml-IN)
Marathi (mr-IN)
Norwegian (no-NO)
Odia (od-IN)
Polish (pl-PL)
Portuguese (Brazil) (pt-BR)
Portuguese (Portugal) (pt-PT)
Punjabi (pa-IN)
Romanian (ro-RO)
Russian (ru-RU)
Slovak (sk-SK)
Spanish (Mexico) (es-MX)
Spanish (Spain) (es-ES)
Spanish (US) (es-US)
Swedish (sv-SE)
Tamil (ta-IN)
Telugu (te-IN)
Turkish (tr-TR)
Ukrainian (uk-UA)
Vietnamese (vi-VN)

ElevenLabs

Scribe v2 Realtime

scribe_v2_realtime · Official docs ElevenLabs calls Scribe v2 Realtime “our fastest and most accurate live speech recognition model”.

Metric	Value	Source
Latency	~150 ms, excluding application and network latency	ElevenLabs
Languages (provider)	90+, with automatic language recognition	ElevenLabs
Features	Voice activity detection, manual commit control, word-level timestamps, entity detection (65 entity types)	ElevenLabs

Languages available in SubVerse AI (50)

Arabic (Kuwait) (ar-KW)
Arabic (Saudi Arabia) (ar-SA)
Arabic (UAE) (ar-AE)
Bengali (bn-IN)
Bulgarian (bg-BG)
Chinese, Mandarin (Traditional, Taiwan) (cmn-Hant-TW)
Croatian (hr-HR)
Czech (cs-CZ)
Danish (da-DK)
Dutch (nl-NL)
English (Australia) (en-AU)
English (Canada) (en-CA)
English (India) (en-IN)
English (UK) (en-GB)
English (US) (en-US)
Filipino (fil-PH)
Finnish (fi-FI)
French (Canada) (fr-CA)
French (France) (fr-FR)
German (de-DE)
Greek (el-GR)
Gujarati (gu-IN)
Hindi (hi-IN)
Hungarian (hu-HU)
Indonesian (id-ID)
Italian (it-IT)
Japanese (ja-JP)
Kannada (kn-IN)
Korean (ko-KR)
Malay (ms-MY)
Malayalam (ml-IN)
Marathi (mr-IN)
Norwegian (no-NO)
Odia (od-IN)
Polish (pl-PL)
Portuguese (Brazil) (pt-BR)
Portuguese (Portugal) (pt-PT)
Punjabi (pa-IN)
Romanian (ro-RO)
Russian (ru-RU)
Slovak (sk-SK)
Spanish (Mexico) (es-MX)
Spanish (Spain) (es-ES)
Spanish (US) (es-US)
Swedish (sv-SE)
Tamil (ta-IN)
Telugu (te-IN)
Turkish (tr-TR)
Ukrainian (uk-UA)
Vietnamese (vi-VN)

Sarvam

Saaras v3

saaras:v3 · Official docs Saaras v3 is Sarvam’s speech recognition model for Indian languages. It supports the 22 official Indian languages plus English, and was trained on more than 1 million hours of audio.

Metric	Value	Source
Word error rate, IndicVoices	~19%, down from ~22% for Saaras v2.5	Sarvam
Word error rate, IndicVoices (10 most popular languages)	19.31%	Sarvam
Features	Automatic language detection, speaker diarization, 8 kHz telephony optimization	Sarvam

Sarvam now lists saaras:v4 as its default, recommended model. Saaras v3 “remains available”.

Languages available in SubVerse AI (11)

Bengali (bn-IN)
English (India) (en-IN)
Gujarati (gu-IN)
Hindi (hi-IN)
Kannada (kn-IN)
Malayalam (ml-IN)
Marathi (mr-IN)
Odia (od-IN)
Punjabi (pa-IN)
Tamil (ta-IN)
Telugu (te-IN)

Saarika v2.5

saarika:v2.5 · Official docs Saarika v2.5 transcribes speech in the language spoken. It handles telephony (8 kHz), multi-speaker and code-mixed audio.

Sarvam says “Saarika v2.5 will be deprecated soon” and recommends Saaras instead.

Metric (VISTAAR benchmark)	Value	Source
Word error rate, all 11 languages	18.32%	Sarvam
Word error rate, English	8.26%	Sarvam
Word error rate, Hindi	11.81%	Sarvam
Character error rate, all 11 languages	4.96%	Sarvam

Languages available in SubVerse AI (11)

Bengali (bn-IN)
English (India) (en-IN)
Gujarati (gu-IN)
Hindi (hi-IN)
Kannada (kn-IN)
Malayalam (ml-IN)
Marathi (mr-IN)
Odia (od-IN)
Punjabi (pa-IN)
Tamil (ta-IN)
Telugu (te-IN)

Providers measure accuracy on different datasets and under different conditions, so word error rates from different providers are not directly comparable. Test candidate models with audio from your own use case.

Overview

Model Types

Reference

At a glance

Deepgram

Nova 3

Flux (English)

Flux (Multilingual)

OpenAI

GPT Live Transcribe

GPT Realtime Whisper

Google

Telephony

Chirp 3

ElevenLabs

Scribe v2 Realtime

Sarvam

Saaras v3

Saarika v2.5

​At a glance

​Deepgram

​Nova 3

​Flux (English)

​Flux (Multilingual)

​OpenAI

​GPT Live Transcribe

​GPT Realtime Whisper

​Google

​Telephony

​Chirp 3

​ElevenLabs

​Scribe v2 Realtime

​Sarvam

​Saaras v3

​Saarika v2.5

At a glance

Deepgram

Nova 3

Flux (English)

Flux (Multilingual)

OpenAI

GPT Live Transcribe

GPT Realtime Whisper

Google

Telephony

Chirp 3

ElevenLabs

Scribe v2 Realtime

Sarvam

Saaras v3

Saarika v2.5