AI Transcription¶
Overview¶
AI Transcription turns call recordings and voicemails into text. For each transcribed recording, QVOICE Platform stores:
- the full transcript, separated by speaker
- a short summary
- an optional sentiment score
Recordings are transcribed with one of two providers, chosen per account:
- Deepgram — transcript, summary and optional sentiment; the language is detected automatically.
- Google Speech-to-Text — transcript only, in a language chosen in the settings or detected automatically. Use it for languages Deepgram does not cover, such as Sinhala.
The same speech-to-text service also powers live transcription during calls in the web client, and meeting transcripts.
Where to Find It¶
Config → Integrations → AI Transcription → Transcription
Functionality¶
Settings¶
Provider selects Deepgram or Google Speech-to-Text. The other fields depend on it.
Deepgram¶
| Field | Description |
|---|---|
| API Key | The Deepgram API key. |
| Model | Nova 3 (default), Nova 2, Nova, Whisper Large, Whisper Medium, Whisper Small or Whisper. |
| Sentiment | Enables sentiment analysis of recordings. |
The spoken language is detected automatically; no language needs to be set.
Google Speech-to-Text¶
| Field | Description |
|---|---|
| Service Account Key (JSON) | The JSON key of a Google Cloud service account. The account's project must have the Cloud Speech-to-Text API enabled, and the service account needs write access to the bucket below. Once saved, the key is never shown again: the screen shows the service account's email instead, and leaving the field empty on a later save keeps the stored key. |
| Cloud Storage Bucket | A bucket in the same Google Cloud project. Each recording is uploaded there for transcription and deleted right after. |
| Language | Sinhala (si-LK) (default), Tamil (ta-IN), English (en-US), English (en-IN) or Auto-detect. With Auto-detect, Google picks the most prevalent language of each recording and transcribes the whole recording in it; a call that switches between languages is not transcribed phrase by phrase in each one. |
| Region | Where Google processes the audio: Europe (europe-west4) (default), United States (us-central1) or Asia (asia-southeast1). |
| One speaker per channel | Transcribes each audio channel separately. Enable it when recordings are stereo with each party on its own channel. |
Recording Translation (Google only):
| Field | Description |
|---|---|
| Translate Transcripts | Lets users translate a recording's transcript in the player. Google Cloud Translation translates it with the service account above; its project needs the Cloud Translation API enabled and the service account the Cloud Translation API User role. Google bills it by translated character. |
| Default Language | The language the player offers first. Any language Google supports can still be picked. |
| Auto Translate | Translates every automatic transcript into the default language as soon as it is ready. Google bills each one, whether or not anyone opens the recording. |
| Check | Translates a test word with the stored key and shows whether it works, or which API or role is missing. |
To prepare the Google Cloud project (APIs, bucket, service account and roles), follow Set Up Google Speech-to-Text; translation is Step 7.
Google transcription uses the Chirp 2 model. It produces the transcript only: there is no summary and no sentiment with this provider.
Both providers¶
| Field | Description |
|---|---|
| Auto Transcribe Recordings | Transcribes every new call recording automatically. When it is off, recordings are transcribed on demand. |
Save enables transcription for the account. Remove deletes the settings and disables it.
Live call transcription¶
Live transcription while a call is in progress is configured per account in Config → Accounts → (account) → Calls:
- Provider: Deepgram or Google Speech-to-Text.
- Mode:
- Automatic: all calls are transcribed.
- Manual: the user turns transcription on per call.
- Language:
- With Deepgram: English, Spanish, Portuguese, French, German, Italian, or multi-language.
- With Google: a regional language code.
Live transcription runs in the web client: only calls the agent answers there are transcribed live. External systems can follow the live transcript of a call through the Live Call Transcription WebSocket.
Where users see the results¶
| Where | What |
|---|---|
| Admin portal: Reports → Call Recordings (play a recording) | Sentiment and score, Summary (Auto-generated) with copy and .txt download, full transcript, and a button to transcribe on demand |
| Admin portal: Reports → Call Recordings and the Transcript Search report, with Recording Translation on | A Translate… list on the transcript; the translation is saved with the recording. With Auto Translate, Transcript Search also finds calls by their translation |
| Admin portal: Call Recordings export | "Download Selected Call Recordings with Transcript" adds the transcripts to the ZIP |
| Web client (agents and users) | Transcript, summary and sentiment in the call detail and recording view; voicemail transcripts; live transcript during calls; meeting transcripts and summaries |
| External systems | The live transcript of a call in progress, through the Live Call Transcription WebSocket |
| CRM | Voicemails linked to a case get their transcript on the case; manual follow-ups on cases are transcribed and summarised |
How It Works¶
flowchart LR
R[Call ends and the recording is stored] --> A{Auto transcribe?}
A -- yes --> D[Account's provider: Deepgram or Google]
A -- no --> M[User clicks Transcribe / opens the recording]
M --> D
D --> S[Transcript saved with the recording, plus summary and sentiment with Deepgram]
S --> P[Portal / web client / CRM]
- Transcription runs after the call, on the stored recording. Only recorded calls can be transcribed.
- In the web client, opening a recording that has no transcript yet requests one automatically.
- Transcribing a recording again replaces the previous result.
- Summaries and sentiment come from Deepgram. They do not use the OpenAI / Gemini keys in GenAI Integrations.
- With Google Speech-to-Text, a recording is copied to the account's bucket only while it is being transcribed.
Configuration¶
| Requirement | Who | Details |
|---|---|---|
| Call recording | Account admin | Calls must be recorded to be transcribed. |
| Deepgram credentials | Customer or platform | Enter the customer's own Deepgram key on this screen, or use the platform's default service if the operator provides one. Contact QVOICE support to confirm. |
| Google credentials | Customer | A Google Cloud project with the Cloud Speech-to-Text API enabled, a Cloud Storage bucket, and a service account key with access to both. Google bills the customer's project. |
| Who can edit | Reseller administrators and platform super-administrators | Administrators of a regular account do not see this screen. |
| Viewing transcripts | Users | The account must have the transcription feature enabled. |
Use Cases¶
Quality assurance without listening to every call. A contact centre turns on Auto Transcribe Recordings and Sentiment. QA staff open Reports → Call Recordings, read the auto-generated summary, and listen to the full recording only for calls with negative sentiment.
Compliance and dispute handling. A customer disputes what was agreed on a sales call. The supervisor downloads the recording together with its transcript from the Call Recordings export and attaches both to the case.
Agent after-call work. Agents open the call detail in the web client right after the call. They copy the summary into their notes instead of writing it from memory.
Voicemail triage. Voicemails left for a support line are transcribed. When a voicemail is linked to a CRM case, the transcript is on the case, so the team can act on it without playing the audio.
Limitations¶
- Only recorded calls and voicemails are transcribed. Automatic transcription covers call recordings; voicemails are transcribed when opened or linked to a case.
- Each transcription is a request to the provider, billed to the credentials in use (the customer's or the platform's).
- Summary and sentiment quality and language coverage depend on Deepgram. Google Speech-to-Text produces no summary or sentiment.
- The Google settings apply to recordings and voicemails only. Live transcription during calls with Google is set separately, in the account's Calls settings.