Gong alternative: sales call analysis on your own server, from recording to CRM note

Self-hosted conversation intelligence for sales calls

A company can analyse its sales calls with open software on its own server instead of Gong: a self-hosted meeting bot such as Vexa, or the recordings the company already makes, provides the audio; WhisperX writes the transcript and labels who is speaking; a language model such as gpt-oss-120b writes the summary, the objections and the next steps; and a short script files the note in the CRM. For a B2B sales team the computing load is small, talk ratio and monologue length are counted from the timestamps, and recordings, transcripts and analyses stay in the company network.

This comparison is for sales directors and sales operations in companies that sell to other companies, with teams of a few to a few hundred people, and for the IT team that runs the server. A contact centre with hundreds of agents on the phone at once is a different task, covered in our guide to contact-centre AI on-premise. For internal meetings, our guide to meeting minutes with a local AI applies.

Conversation intelligence is the software category Gong belongs to: it records customer calls and evaluates them. Transcription turns speech into text with timestamps. Speaker labelling, also called diarisation, marks which person said which sentence. A language model (LLM) is the AI that reads the transcript and writes the summary. The CRM is the customer database in which the sales team keeps its deals.

Gong's features and data handling, as Gong describes them

Gong's conversation intelligence page, read on 30 October 2026, says the product "automatically records and transcribes every sales call, meeting, and email". Its AI analyses conversations for "key topics, objections, competitor mentions, and buying signals", automates "CRM updates, summaries, and next steps", helps managers "Coach reps with actionable, data-backed insights from real conversations" and can "Surface deal risks and buying signals in real time".

On data, the same page states "Your data is stored securely in enterprise-grade cloud infrastructure" and that retention policies are configurable. Gong's security page describes "segregated customer data environments hosted on AWS", encryption "in transit and at rest", bring-your-own-key encryption, "consent management (with support for 70+ languages)" and "tools for recording consent", and states that Gong "is certified with the EU-U.S. Data Privacy Framework". Neither page names a hosting region; Gong's help centre (updated 15 September 2025, read on 31 October 2026) states that new customers "can decide to store their data in data centers in either the US or EU", with the US as the default.

Gong covers more than a local set-up does at the start, from email analysis to coaching and deal-risk warnings. A local chain covers the core, the call itself: what was said, by whom, what the customer objected to and what was agreed. The audio and the transcripts then do not leave the company, and the analysis follows your own list of topics.

Five steps from call to CRM note with open tools

Vexa-ai repository on GitHub, whisperx on PyPI, Hugging Face cards of openai, pyannote and nvidia, read 30 October 2026; gpt-oss licence as recorded in our model guides.

Record. Vexa describes itself as an "Open-source meeting transcription API" for Google Meet, Teams and Zoom, all three marked as production in its README, under Apache 2.0. Vexa sends a bot into the video call, where it appears as a participant. Vexa's own stack runs without a GPU; transcription is a separate unit that you host yourself or replace with Vexa's hosted service, so the set-up decides whether audio leaves the building. Where the phone system or the meeting platform already stores recordings, a folder on the server that receives them is enough, and the recordings then also exist on that platform. For conversations in person, our guide to an AI voice recorder without the cloud describes the route.

Transcribe. WhisperX, version 3.8.6 of 25 May 2026 under BSD-2-Clause, runs Whisper with word timestamps and includes alignment models for German, English, French, Spanish and Italian by default; the German, French, Spanish and Italian defaults are torchaudio's VoxPopuli models under CC BY-NC 4.0, a non-commercial licence, so we set alignment models whose licence allows commercial use. On public German test sets Whisper large-v3 makes about one error in twenty words, as our guide to German speech recognition shows; product names and industry terms deserve a test with your own calls.

Label speakers. WhisperX uses pyannote's speaker-diarization-community-1, under CC-BY-4.0; the model is gated on Hugging Face, and a local copy "works without internet connection". NVIDIA's Nemotron 3 Diarization handles up to eight speakers under the OpenMDW License 1.1. Vexa attributes speakers itself; its README notes that under heavy crosstalk it "publishes an empty speaker rather than guessing", for about 4 to 7 per cent of rows. Our guide to speaker diarisation on-premise compares the models.

Analyse. The language model reads the labelled transcript with a fixed instruction and returns a fixed structure: a summary in five lines, the customer's objections with a quote and a timestamp, competitors and products mentioned, agreed next steps with owners and dates, and open questions. The list of topics to watch, such as price, delivery terms, a competitor or a product line, comes from your sales team.

File. A script or a workflow tool such as n8n, whose licence our guide to n8n with a local LLM explains, writes the summary and next steps as a note on the deal in the CRM. Only that text goes to the CRM; if the CRM is a cloud service, the note is stored there, while audio and full transcript stay on your server.

You can order the server alone, the setup of Vexa, WhisperX and the language model alone, or both together: the server delivered with everything installed and configured; write to us and we suggest the best solution for your tasks and budget.

What to measure, and what to leave out

Some figures need no AI at all. From the labelled transcript a short script counts each speaker's share of the talking time (talk ratio), the longest monologue, the number of questions asked by the rep and the share of the call spent on the customer's topics. These numbers are repeatable and as exact as the speaker labels. The language model adds what needs reading: what the objection was, whether a next step was agreed, which competitor came up.

The AI Act draws two lines here. Article 5(1)(f) prohibits AI systems "to infer emotions of a natural person in the areas of workplace", except for medical or safety reasons, and that prohibition has applied since 2 February 2025. A sentiment or stress score derived from the sales rep's voice belongs in that zone, so leave it out; the same score from the customer's voice would be an emotion recognition system, which Annex III, point 1(c), lists as high-risk. Annex III, point 4(b), lists as high-risk AI systems intended "to monitor and evaluate the performance and behaviour of persons in such relationships", meaning work relationships. A scorecard that ranks each rep comes close to that point; under the amended Article 113 these rules apply from 2 December 2027, and Article 26(7) then requires employers to "inform workers' representatives and the affected workers" before such a system is used. Call summaries and deal notes for the rep's own follow-up are a different use, and that is the use we set up: summaries, objections and next steps per call, with no emotion analysis and no scores or rankings of individual reps. Our article on the EU AI Act for companies explains the deployer duties, and your legal counsel confirms how your use is classified.

Recording consent, the GDPR and the works council

A recorded sales call carries personal data of the customer's staff and of your rep. Article 13(1) of the GDPR requires the controller to inform people "at the time when personal data are obtained", which for a call means at its start: in the invitation, by the bot's name in the participant list and by a sentence from the rep that asks whether everyone agrees to the recording; if someone does not, the rep removes the bot from the call. The legal basis may be consent or "the legitimate interests pursued by the controller" under Article 6(1)(f); which applies is for your data-protection officer. Article 5(1)(c) asks for data "limited to what is necessary", so decide which calls are recorded, delete the audio once the transcript is checked, and keep transcripts only as long as the deal needs them.

Recording conversations is also governed by national law, which differs between EU countries and is not covered here; your legal counsel confirms what applies to calls with customers in each country. For your own staff, workers' representatives often have a say in recording and evaluation under national law; our guide on the works council and an internal AI assistant covers the EU framework, and a DPIA for an internal LLM is the usual first document. This section is general information as of October 2026, not legal advice.

Sizing by call hours per day

A sales team is sized by its call hours per day. Take the number of reps, multiply by their customer-call hours per day, and plan the processing overnight or within an hour of each call. Three published figures give the scale. The Open ASR Leaderboard paper (arXiv 2510.06961v3) measured Whisper large-v3 at an RTFx of 145.5 on an NVIDIA A100, which means 145.5 seconds of audio per second of computing. NVIDIA's card for Nemotron 3 Diarization gives an RTFx of 1,340 on an RTX PRO 5000 Blackwell (offline profile, batch 1, BF16). And in MLPerf® Inference v6.1, eight RTX PRO 6000 Blackwell Server Edition generated 15,256 tokens per second with gpt-oss-120b in the offline scenario, which is 1,907 per card by our division.

Our arithmetic: Whisper large-v3 at RTFx 145.5 on one A100-SXM4-80GB (Open ASR Leaderboard paper); Nemotron 3 Diarization at RTFx 1,340 on one RTX PRO 5000 Blackwell (NVIDIA model card); analysis at 800 output tokens per 30-minute call (our example) and 1,907 tokens per second per card, our division of MLPerf® Inference v6.1, data-centre suite, closed division, available systems, gpt-oss-120b, offline scenario, entry 6.1-0050 (HPE ProLiant Compute DL380a Gen12, eight RTX PRO 6000 Blackwell Server Edition), retrieved from mlcommons.org on 24 October 2026, result verified by MLCommons Association; the per-card figure is not an MLPerf metric. The MLPerf name and logo are registered and unregistered trademarks of MLCommons Association in the United States and other countries. All rights reserved. Unauthorized use strictly prohibited. See www.mlcommons.org for more information.

The figures come from different GPUs and from the benchmark's own request mix; a long German transcript as input lowers the token rate, and no publisher states Whisper's speed on an RTX PRO card. Read the table as an order of magnitude: even a large sales organisation needs a few GPU hours a day, so the choice of hardware follows the size of the language model, the number of teams sharing the machine and where it stands. A trial with a week of your own calls confirms it.

Which hardware for a sales team

usable memory by our sizing rule (90 per cent of driver-visible memory, less 3 GiB per card): 25.7 GiB on 32 GB, 83.0 GiB on 96 GB; Whisper large about 10 GB (OpenAI), gpt-oss-20b 12.8 GiB and gpt-oss-120b 60.8 GiB (about 65 GB) of weights as in our model guides, 1.125 GiB of 16-bit cache per 32K-token analysis with gpt-oss-120b; wall power by our estimate.

Entry. One NVIDIA RTX PRO 4500 Blackwell with 32 GB holds Whisper large (about 9.3 GiB) and gpt-oss-20b (12.8 GiB) within its 25.7 GiB, with 3.6 GiB left, enough for a few analyses at a time overnight. It suits a pilot with one team.

Department. A workstation with one RTX PRO 6000 Blackwell Workstation Edition (96 GB, 600 W) carries the larger gpt-oss-120b next to Whisper with 12.9 GiB left by our rule, room for 11 analyses with 32K tokens of context at once. By our estimate it draws about 1.1 kW at the wall.

Server. The entry server on our AI servers page has two RTX PRO 6000 Server Edition cards, two AMD EPYC processors, 384 GB of DDR5 ECC memory and two NVMe drives. One card runs transcription and speaker labelling, the other gpt-oss-120b with 22.2 GiB left for 19 analyses with 32K tokens of context at once, as in our guide to meeting minutes. The meeting bot has to be reachable whenever a call starts, so this machine belongs in a server room, where it draws about 1.8 kW by our estimate on one 16 A feed. The same server can run the company's internal assistant or the meeting minutes.

Large. A node with four Server Edition cards, about 3.4 kW by our estimate, fits when sales, service and a contact centre share one platform; our contact-centre guide sizes the speech and LLM cards for the contact-centre share.

Send us the number of reps, their call hours per day, the languages and your CRM, and we propose the configuration that fits.

Vexa, WhisperX and vLLM in operation

1. Vexa. Self-hosting runs with Docker Compose (Docker engine v26 or later), with Ubuntu 24.04 as the production target, a single-container image and a Helm chart. The README names version 0.12; in the open-source stack WebSocket streaming is not wired yet, and transcripts are polled. Host the transcription unit from deploy/transcription yourself if audio must stay in house.
2. WhisperX. Diarisation needs a Hugging Face read token passed with --hf_token and acceptance of the model's user agreement for the first download; afterwards a local copy works offline. GPU use requires the CUDA toolkit 12.8, which RTX PRO Blackwell cards need anyway.
3. The language model. vLLM's --api-key protects only the paths under /v1, /v2, /inference and /cohere; /metrics, /health and /invocations answer without the key, and /invocations gives the same inference as /v1 (vLLM security guide, read 30 October 2026). Keep the service on the internal network behind an LLM gateway with keys and a query log.
4. Storage. Separate audio, transcripts and analyses by team with access rights, delete audio on a schedule, and keep the bot's account and the CRM token out of personal accounts.

FAQ