Speech-to-Text Integrations
Speech-to-text converts what callers say into text that your LLM can process. Transcription accuracy and latency directly affect how natural a conversation feels. Bolna supports streaming STT providers optimized for telephony audio, including specialized models for Indian languages.
AssemblyAI
AssemblyAI Universal-3.6 Pro transcribes callers for your Bolna voice agent in 32 languages, with code-switching and keyword plus context biasing.
Azure Speech-to-Text
Azure Speech-to-Text transcribes callers in real time for a Bolna voice agent, with speaker diarization, language identification and phrase lists of up to 2,000 terms.
Deepgram
Deepgram is Bolna's default English transcriber, turning caller speech into text with Nova-3, Nova-2 or Flux streaming models on live phone calls.
ElevenLabs Scribe (STT)
ElevenLabs Scribe v2 Realtime transcribes callers for a Bolna voice agent in about 150ms, across 90 languages, with word-level timestamps and auto language detection.
Google Gemini (STT)
Google's Gemini 3.5 Transcribe gives a Bolna voice agent streaming speech-to-text in 19 language codes, including Hindi, Tamil and Bengali, on phone and web calls.
Gladia
Gladia's Solaria model gives a Bolna voice agent real-time speech-to-text in 100+ languages with code-switching, custom vocabulary and under 300ms of latency.
OpenAI (STT)
OpenAI's GPT Realtime Whisper streams speech-to-text for a Bolna voice agent, with built-in voice activity detection, noise reduction and a delay setting for accuracy.
Sarvam AI (STT)
Sarvam's Saaras models transcribe Indian callers for your Bolna voice agent in Hindi, English and 10 more Indian languages, including code-mixed Hinglish.
Smallest AI (Pulse)
Smallest AI's Pulse model transcribes callers for your Bolna voice agent with sub-100ms first-token latency, keyword weights and speaker labels.
Soniox
Soniox stt-rt-v5 transcribes callers for your Bolna voice agent with one multilingual model that auto-detects and code-switches, including Hinglish.
Speech-to-text providers on Bolna compared
| Provider | Best for | Languages | Notes |
|---|---|---|---|
| Sarvam (saaras:v4) | Indian languages | 11 Indian languages plus English on Bolna | Handles regional accents and code-mixed speech; included in the flat rate |
| Deepgram Nova-3 | English | English | Fastest and most accurate for English, with sub-150ms transcription latency; Bolna's default English transcriber |
| Deepgram Nova-2 | Many languages; cost-sensitive volume | 30+ languages | Set the language explicitly; included in the flat rate |
| Deepgram Flux | Lowest latency | English, plus a multilingual variant | Turn detection inside the model allows instant barge-in |
| Soniox (stt-rt-v5) | Code-switching, such as Hinglish | English and 9 Indian languages, plus multilingual auto-detect | One model detects and switches language mid-sentence |
| AssemblyAI (universal-3-6-pro) | Short answers and proper nouns | 32 languages, including Hindi, Marathi and Urdu | Code-switches across its languages; accepts keywords and free-form context |
| Smallest AI (Pulse) | Ultra-low latency | 31 documented languages, including Hindi, Tamil, Telugu, Kannada and Malayalam | Sub-100ms first-token latency; set model to pulse explicitly |
| ElevenLabs Scribe (scribe_v2_realtime) | Multilingual agents on the flat rate | 90 languages, with automatic language detection | About 150ms latency; word-level timestamps; included in the flat rate |
| Gladia (Solaria) | Code-switching, including Hinglish | 100+ languages | Sub-300ms latency; custom vocabulary; audio enhancer on phone calls |
| Azure Speech-to-Text | Enterprise deployments | Multilingual, with language identification | Speaker diarization; phrase lists of up to 2,000 terms; included in the flat rate |
| Gemini 3.5 Transcribe | Indian languages on Google | 19 language codes, including Hindi, Tamil, Telugu and Bengali | Works on 8 kHz phone and 16 kHz web audio |
| OpenAI GPT Realtime Whisper | Accented or noisy audio | Many languages, set by ISO 639-1 code | Built-in VAD; a delay setting trades latency for accuracy; optional noise reduction |
For the full guide, see Choosing a speech-to-text provider in the Bolna docs.
Frequently asked questions
Soniox is built for it: choose its multilingual option and one model transcribes English and Hindi in the same sentence. Sarvam also handles code-mixed English and Indian-language speech, and AssemblyAI universal-3-6-pro code-switches across its 32 languages.
Deepgram Flux has the lowest latency, because turn detection runs inside the model. Smallest AI Pulse has sub-100ms first-token latency, and Deepgram Nova-3 transcribes English in under 150ms.
Set it explicitly. Auto-detection adds latency (100 to 200ms on Deepgram) and can pick the wrong language. The exception is callers who mix languages, where Soniox in multilingual mode is the better fit.
Deepgram nova-2 and nova-3, Azure, Sarvam saaras:v4, saaras:v3, saaras:v2.5 and saarika:v2.5, ElevenLabs scribe_v2_realtime, and Soniox stt-rt-v5. Other transcribers are billed at usage-based rates.
The endpointing value is too low for those callers. Raise it from the 250ms default to 350 to 500ms for non-native speakers, elderly callers or anyone who pauses mid-sentence.
Last updated