Skip to main content

Transcription

Settings → Transcription lists the recognisers: the services that turn audio into text, for finished conversations and, for the Prompter, while a conversation is happening.

A conversation is transcribed when you ask for it in the Recordings window, or by itself if Process conversations automatically is on in Processing. A recogniser on your own machine costs nothing to run; one in the cloud charges by the minute of audio.

Recognisers​

A recogniser is a speech-to-text service the phone sends audio to. Press Add to add one; the Test button of its card checks that the service really answers. Each is listed with its name and, under it, the model and the address of its service. The table lists every kind of service the phone supports, as Kind on the card offers them, with the two addresses and the models a new recogniser of that kind is given; the last column shows the five set up in the picture. The address for the prompter is usually a WebSocket one: wss:// for a service in the cloud, ws:// for a server in your own network; NVIDIA Riva is given as host:port.

KindAddress for transcriptsAddress for the prompterModelsIn the picture
OpenAI-compatible (Whisper, OpenAI)https://api.openai.com/v1wss://api.openai.comgpt-4o-transcribe, live gpt-live-transcribeOpenAI · gpt-4o-transcribe
Deepgramhttps://api.deepgram.com/v1wss://api.deepgram.comnova-3Deepgram · nova-2
ElevenLabs (Scribe)https://api.elevenlabs.io/v1wss://api.elevenlabs.ioscribe_v1, live scribe_v2_realtimeElevenLabs · scribe_v2
Speechmaticshttps://asr.api.speechmatics.com/v2wss://global.rt.speechmatics.com, or eu., us., au. in place of global. for a regionenhanced—
xAI (Grok)https://api.x.ai/v1wss://api.x.ai(empty: the service default), live grok-voice-transcribe-2.0X.ai · (empty), live grok-voice-transcribe-2.0
Sonioxhttps://api.soniox.comwss://stt-rt.soniox.com(empty: the service default), live stt-rt-v5—
AssemblyAIhttps://api.assemblyai.comwss://streaming.assemblyai.com(empty: the service default)—
Vosk (on your own machine)your server, for example ws://localhost:2700(the same server)(set on the server)Vosk · ws://localhost:2700
WhisperLive (on your own machine)your server, for example ws://localhost:9090(the same server)live small—
NVIDIA Riva (on your own machine)your server, for example localhost:50051(the same server)(set on the server)—

The two marks at the right of a row say what the recogniser is the default for. The clock is lit on the default for transcripts — X.ai in the picture — which is used when you do not choose another. The lightning is lit on the default for the prompter — Vosk in the picture. You can keep several recognisers; the drop-down above a transcript in the Recordings window lists the transcripts made by each one.

The recogniser's card​

Pressing a recogniser opens its card.

FieldWhat it is
NameThe name in the lists.
KindThe kind of service, which decides how the phone talks to it: OpenAI-compatible (Whisper, OpenAI), Deepgram, ElevenLabs (Scribe), xAI (Grok), AssemblyAI, Soniox, Speechmatics, and three that run on your own machine — Vosk, WhisperLive and NVIDIA Riva.
Address for transcriptsWhere finished conversations are sent.
Address for the prompterWhere live audio is sent while a conversation is happening. It is usually a WebSocket address: wss:// for a service in the cloud, such as wss://api.x.ai in the picture, ws:// for a server in your own network. Empty is worked out from the address beside it.
KeyThe service's key. It is kept in this computer's keyring, never in a settings file.
TestAsks the service and says what it answered, for example Answered, and offers 3 model(s).
Model for transcripts and Model for the prompterThe model, exactly as the service names it. Empty sends no model name, and the service uses its own default; where the vendor publishes one, the card names it. A field is not shown for a kind that has no choice — as Model for transcripts for xAI in the picture, where Model for the prompter is set to grok-voice-transcribe-2.0.
Default for transcriptsMakes this the recogniser used when you do not choose another.
Default for the prompterMakes this the recogniser a new assistant of the prompter listens with.
EnabledOff, the recogniser stays in the list and is not used.

Advanced settings opens the rest of the card. The values that matter most:

FieldWhat it does
RegionThe region of the service, for one that has several.
Send the two sides separatelyA call is recorded with the two people on two channels, which is what tells the recogniser who said what. Turn it off for a server that says it can do this and cannot.
Ask who is speakingNames the people inside one channel, where several speak on it.
Write numbers as figuresSums, dates and phone numbers come back as they are written rather than spelt out.
Upload limit, Length limitThe largest file, in bytes, and the longest recording, in seconds, this phone will send.
Requests at onceHow many requests may be in flight at the same time.
End a reply after, Join short replies within, Turn gapFor the prompter: how long without new words ends a reply, how long a short reply waits for the next one to be joined to it, and how long a silence ends a turn where the recogniser marks none. In milliseconds.
LanguageA two-letter language code as in ISO 639-1 (en, de, es, fr, sr…). Leave it empty and the recogniser decides — that is right unless your calls are in a language it keeps mishearing.
ExtrasOne name = value a line, passed to the service as it is. Leave it empty unless the server documents something.
Wait, minutesHow long to wait for a transcript. Empty works it out from the length of the recording.
Price a minuteWhat a minute of live audio costs, from the service's price list. The prompter shows what a session has cost and stops at its monthly ceiling.

Live recognition for the prompter​

The Prompter needs a recogniser that listens while somebody is talking, over a stream rather than with a finished file. These kinds can: Deepgram, ElevenLabs (Scribe), xAI (Grok), OpenAI-compatible (with OpenAI's realtime transcription), AssemblyAI, Soniox and Speechmatics in the cloud, and Vosk, WhisperLive and NVIDIA Riva on your own machine. A recogniser on your own machine keeps the other party's voice in the building and charges nothing.

To use one: open its card and check the Address for the prompter — usually a WebSocket address, such as wss://api.x.ai for X.ai, wss://api.deepgram.com for Deepgram or ws://localhost:2700 for Vosk; left empty, it is worked out from the address for transcripts. Then choose the Model for the prompter: the live models are often different from the ones for files — grok-voice-transcribe-2.0 for X.ai, as on the card in the picture, scribe_v2_realtime for ElevenLabs, gpt-live-transcribe for OpenAI, stt-rt-v5 for Soniox. Press Test. Tick Default for the prompter to make it the one new assistants listen with.

Which model to choose​

The table lists the speech-to-text models of every kind in the Kind list. The models in bold are the ones set up in the picture; for the X.ai recogniser the model for transcripts is empty, so the service default, grok-voice-transcribe-2.0, is the one used, and the model for the prompter is set to the same model. For says what a model is made for: finished recordings (transcripts), live speech for the prompter (prompter), or both.

Kind and addressModelForWhat it is for
OpenAI-compatible (Whisper, OpenAI)
https://api.openai.com/v1
gpt-transcribetranscriptsThe model OpenAI recommends for recorded speech in its original language.
gpt-4o-transcribebothGeneral-purpose transcription. A new recogniser of this kind is given it.
gpt-4o-mini-transcribebothA lighter, cheaper variant of the above.
gpt-4o-transcribe-diarizetranscriptsLabels who speaks when. Use it only if you need that.
whisper-1transcriptsThe older Whisper model, kept for special uses such as word timestamps and subtitles.
gpt-live-transcribeprompterOpenAI's live model: the words come as they are spoken. The phone offers it for the prompter.
Deepgram
https://api.deepgram.com/v1
nova-3bothDeepgram's best general-purpose model, for meetings, noisy and multilingual audio. A new recogniser of this kind is given it.
nova-2bothThe previous generation; keep it for languages nova-3 does not support yet.
nova-2-phonecallbothnova-2 tuned for the narrow sound of a phone line. English.
flux-general-enprompterMade for conversation: it hears when somebody has finished speaking. English.
flux-general-multiprompterThe same in ten languages, and a conversation may switch between them.
enhanced, basetranscriptsOlder tiers; base is for large volumes.
whispertranscriptsWhisper, run by Deepgram.
ElevenLabs (Scribe)
https://api.elevenlabs.io/v1
scribe_v2transcriptsGeneral-purpose transcription in 90+ languages, with speaker separation.
scribe_v2_realtimeprompterThe live version of scribe_v2. The phone offers it for the prompter.
scribe_v2_medicaltranscriptsscribe_v2 tuned for clinical audio.
scribe_v1transcriptsThe first generation; deprecated, use scribe_v2.
Speechmatics
https://asr.api.speechmatics.com/v2
enhancedbothThe most accurate, for a conversation in one language. A new recogniser of this kind is given it.
standardbothFaster and cheaper, a little less accurate.
melia-1transcriptsA conversation in several languages, switching mid-sentence, comes back as one transcript. Recordings only, in the EU and US regions; no custom dictionary or speaker labels yet.
xAI (Grok)
https://api.x.ai/v1
grok-voice-transcribe-2.0bothThe only current model and the default: 38 languages, recognised by itself. Files up to 500 MB; live speech goes to wss://api.x.ai.
grok-voice-transcribe-1.0transcriptsDeprecated: the service sends it to 2.0.
Soniox
https://api.soniox.com
stt-async-v5transcripts60+ languages, with speaker separation.
stt-rt-v5prompterLive, in the same 60+ languages, and hears where a turn ends. The phone offers it for the prompter.
AssemblyAI
https://api.assemblyai.com
universal-3-5-probothThe most accurate model for recordings; 18 languages, and a conversation may switch between them.
universal-2transcripts99 languages, cheaper; AssemblyAI falls back to it for a language universal-3-5-pro does not know.
universal-3-6-proprompterAssemblyAI's newest live model, 32 languages; the service uses it when the model is empty.
universal-streaming-multilingualprompterCheaper live recognition in English, Spanish, German, French, Portuguese and Italian.
universal-streaming-englishprompterCheaper live recognition in English only.
Vosk (on your own machine)
ws://localhost:2700
(set on the server)bothFree and light; runs without a graphics card. The model is the one the server was started with, one per language, for example vosk-model-en-us-0.22 or the small vosk-model-small-en-us-0.15.
WhisperLive (on your own machine)
ws://localhost:9090
smallbothWhisper over a live stream. The size is chosen on the card: tiny, base, small (what the phone offers), medium, large-v3; the larger, the more accurate, and the more it wants a graphics card.
NVIDIA Riva (on your own machine)
localhost:50051
(set on the server)bothNVIDIA's speech server, for a computer with an NVIDIA graphics card. It serves models such as Parakeet and Canary.

Things worth knowing before you choose:

  • Transcripts or prompter. A model made for live speech does not take a finished file, and most models for files cannot listen live. That is why a card has two fields, Model for transcripts and Model for the prompter.
  • File size. OpenAI takes files up to 25 MB; X.ai up to 500 MB. A long conversation can be larger than a cloud service accepts.
  • Price. Cloud services charge by the minute of audio, and the rates differ by model and change; read them on the service's own page before you switch. A recogniser on your own machine costs nothing to run.
  • Languages. Every service has its own list; check yours, and set the Language code under the recogniser's advanced settings if it guesses wrongly.

The list of a service's models changes often. If a model you want is missing here, the service's own documentation has the current list — OpenAI, Deepgram, ElevenLabs, Speechmatics, X.ai, Soniox, AssemblyAI, Vosk, WhisperLive, NVIDIA Riva — the Model is the name exactly as the service gives it.

Your own models​

A recogniser does not have to be a cloud service. The phone can use any model that is served through the OpenAI-compatible API — the POST /v1/audio/transcriptions interface — whether it runs locally on your computer or on a server of your own. The audio never leaves your premises, nothing is charged by the minute, and there is no limit on volume.

To add one, press Add and give:

  • the address of the server, up to and including /v1, for example http://localhost:8000/v1 for the computer itself or http://asr.local:8080/v1 for a server in your network;
  • the model name exactly as the server lists it, for example openai/whisper-large-v3-turbo.

What can be used​

The usual choice is Whisper, OpenAI's open speech recognition model. It is free to use, understands about a hundred languages, and comes in several sizes: a small model runs on an ordinary computer, the large ones are noticeably more accurate and are best given a graphics card.

ModelNotes
whisper-large-v3The most accurate Whisper. For a server with a GPU.
openai/whisper-large-v3-turboA faster version of large-v3 with a small loss of accuracy.
Systran/faster-whisper-large-v3large-v3 converted for the faster-whisper engine; quicker, and lighter on memory.
medium, small, baseSmaller Whisper models, for a computer without a graphics card.

Whisper is the model that these servers are built around. Some of them can also serve other speech recognition models, such as NVIDIA Parakeet.

Servers that offer the OpenAI-compatible API​

The model has to be run by a server that offers the OpenAI-compatible /v1/audio/transcriptions endpoint. These do:

ServerWhat it is
vLLMA high-performance model server. Serves Whisper at http://localhost:8000/v1 once started.
SpeachesA server for speech models, "Ollama for speech", built on faster-whisper. Loads a model when it is first asked for.
whisper.cppRuns Whisper efficiently on a CPU, including Apple silicon. Its whisper-server is started with --inference-path /v1/audio/transcriptions.
LocalAIA drop-in OpenAI replacement that runs models locally.

Any other server that offers the same endpoint works in the same way. If a server needs a key, enter it as for a cloud service.

Before you rely on a server, make a test recording and look at the transcript in the Recordings window: a conversation in a language the model knows poorly shows it at once.