Transcriber

Talkr platform supports Deepgram, OpenAI, Google, Azure Speech, AssemblyAI, Speechmatics, Cartesia, Gladia, Sarvam, Smallest AI, Hugging Face, and ElevenLabs. You can take a look at the providers documentation of which language to select for your language requirements.

For locally deployed or self-hosted STT models, Talkr also supports Speaches, an OpenAI API-compatible server for streaming transcription.

Example: Deepgram has their language support documentation at https://developers.deepgram.com/docs/models-languages-overview#nova-3

Background noise suppression

Talkr filters background noise out of the caller's audio in real time before it ever reaches your transcriber — office chatter, traffic, fans, and other steady background sound are suppressed so the transcriber has a cleaner signal to work with. This runs automatically on every call, for every transcriber provider, at no extra cost, and there's nothing to configure.

Choosing a Deepgram model for latency

Deepgram offers two families, and the choice materially changes how fast your agent replies.

Because Flux signals the end of a caller's turn itself, your agent can stop waiting twice — once for a pause timer, then again for the transcript to be finalized. Together those two waits typically account for around half the time a caller spends waiting for a reply, which makes this the single largest speed improvement available to most agents.

New agents use flux-general-multi by default.

Agents you built earlier keep the model you chose — nothing changes underneath a running agent. To switch, pick flux-general-multi in the agent's transcriber settings. If your agent speaks one of the ten supported languages, it's the highest-impact change you can make in one dropdown.

Multi-language / code-switching calls

If your agent needs to auto-detect and transcribe more than one language in the same call (see Voice & Language Profiles), set your Deepgram Language field to multi and pick a model built for it. Deepgram offers two: nova-3-general (Nova-3's general auto-detect) and flux-general-multi (Flux's multilingual/code-switching model). In production testing on real bilingual calls, flux-general-multi recognized short or accented utterances more reliably than nova-3-general, which occasionally missed them entirely rather than transcribing them incorrectly — worth trying if you see an agent seem to "not hear" what was actually said.

Provider-native tuning

The platform-wide Turn Detection settings control how Talkr itself decides a caller has finished speaking. Separately, several transcriber providers expose their own native turn-detection and audio-processing controls beneath that — optional, provider-specific fields that you can leave blank to use the provider's own default. These are advanced settings: they trade transcription speed against accuracy, so change them deliberately and test against real calls rather than guessing at a value.

AssemblyAI

• End of Turn Confidence Threshold, Min Turn Silence, Max Turn Silence — how confident AssemblyAI must be, and how long a pause must last, before finalizing a turn.
• Mode (U3 Pro models only) — a latency/accuracy preset: min_latency, balanced, or max_accuracy.
• Voice Focus (U3 Pro only) — isolates the caller's voice from background noise. Choose near-field for headsets and handsets, far-field for conference rooms and laptop mics, and tune how aggressively background audio is suppressed with Voice Focus Threshold.
• Also available: Language Detection, Format Turns, Speaker Labels, VAD Threshold, Domain, Continuous Partials, Interruption Delay, Previous Context Turns, Word Finalization Max Wait Time, and Formatted Finals — see AssemblyAI's own documentation for what each one changes.

Speechmatics

• End of Utterance Silence Trigger and Max Delay — how much silence ends an utterance, and the longest Speechmatics will wait before finalizing a transcript. Lower values reduce latency at some cost to accuracy.

Soniox

• Endpoint Sensitivity and Max Endpoint Delay — how aggressively Soniox finalizes turns.
• Enable Speaker Diarization, Enable Language Identification, Language Hints Strict, Endpoint Latency Adjustment Level — additional recognition controls.

ElevenLabs (Realtime STT)

VAD Silence Threshold, VAD Threshold, Min Speech Duration, and Min Silence Duration tune ElevenLabs' own voice-activity detection, independent of Talkr's platform-level VAD (see What is VAD?).

Deepgram

Endpointing and Profanity Filter are configurable (previously fixed platform defaults) — Endpointing controls how many milliseconds of silence end a turn. Also available: Diarize, Interim Results, Keywords, Numerals, Punctuate, Redact, Smart Format, Utterance End (ms). On Flux models, Min Confidence sets the minimum confidence Deepgram needs before finalizing a turn.

Google

Nine toggles give fine control over recognition output: automatic punctuation, spoken punctuation, spoken emojis, profanity filtering, word time offsets, word confidence, interim results, voice activity events, and per-channel recognition.

Gladia

Endpointing, Enable VAD, and Maximum Duration Without Endpointing control Gladia's own silence-based turn detection.

Smallest AI

Word Timestamps, Sentence Timestamps, Full Transcript, Numerals, and Diarize control transcript detail; Redact PII and Redact PCI strip sensitive data from the transcript for compliance-sensitive use cases.

Azure Speech

Profanity chooses whether profanity is left raw, masked, or removed from the transcript.

OpenAI & Speaches

Prompt biases transcription toward domain-specific vocabulary — names, acronyms, or jargon your callers are likely to say. Temperature controls sampling randomness during transcription.