Self-hosted

Meeting and call transcription to CRM notes

faster-whisper or WhisperX transcribes and separates speakers on your own GPU, a local model writes the summary and the action items against a fixed schema, and n8n files the result against the right CRM record with a person approving before it is saved.

Save to your workspace

Source: Editorial recipe — no step evidence has been fetched yet
Verified: Evidence not verified
Confidence: Low

01Objective

Turn recorded meetings and calls into accurate transcripts, short summaries and structured CRM notes, without sending client conversations to a transcription service.

Reference deployments using this recipe

Call transcription and CRM notes for a customer-service teamTwenty-five agents writing up calls from memory between calls, badly and late. A self-hosted transcription and summarisation pipeline that drafts the CRM note, with the agent editing and saving it — and with the recording never leaving the company network.Reference deployment · R-05

02Recommended stack

6 components

Role	Component	Notes
Inference	Qwen2.5-14B-Instruct for summaries	Summary, action items and CRM fields, produced against a schema rather than free text.
Inference	Whisper large-v3	Accuracy varies by language and accent far more than the headline numbers suggest — test on your own recordings before quoting a figure to anyone.
Ingestion	faster-whisper	Fast CTranslate2 implementation of Whisper with int8 and float16 options, batched transcription and a VAD filter. The default choice when speaker labels are not required.
Ingestionoptional	WhisperX	Adds word-level alignment and speaker diarisation. Needed the moment a summary has to say who committed to what.
Orchestration	n8n	Watch folder or recording webhook, transcribe, summarise, match the CRM record, queue for approval.
Storage	PostgreSQL	Transcripts, summaries, approval state and retention timestamps.

Architecture and data flow

External data transfer · SOME

confidential content leaves your premises for your own cloud tenancy ("CRM"). You keep control of the account; the provider is a processor, so a DPA and a documented region apply.
Audio and transcripts never leave your network. Only the approved note is written out, and only into your own CRM tenancy.
If the CRM is a SaaS product, that final hop inherits the CRM vendor's terms — state it explicitly in the review.

03Suitable for

Organisation size: 10–500 employees
Data classes: confidential, personal
Constraints: recordings already exist, or you have a lawful basis and a consent process to create them; a CRM with an API; a 24 GB GPU handles transcription and summarisation for a small sales or advisory team; multilingual or accented audio needs testing before anyone is promised accuracy
Industries: Professional services, Financial services, Insurance, Real estate, Technology
Jurisdictions: any

04Hardware

On-premise single 24 GB GPU server
GPU
NVIDIA RTX 4090 24 GB (or NVIDIA L4 24 GB for a rack-mounted, 72 W alternative)
VRAM
24 GB
System RAM
64 GB
Storage
2000 GB
CPU
16-core x86 server CPU (AMD EPYC 7003/9004 or Intel Xeon Scalable)
Form factor
Tower server
Indicative costUS$4,000 – US$9,000
indicative build cost for the complete machine, USD, Aug 2026 — verify with a local supplier
The default box for a 20–60 person firm. Fits a 14B model at 4-bit with roughly 8 GB of KV cache left for concurrent chat, or an 8B model at fp16. NVMe storage sized for the model cache plus a document corpus and its embeddings. Add a UPS and an offsite backup target — this machine holds the whole knowledge base.
Cloud GPU instance — 1 × NVIDIA L4 24 GB
GPU
NVIDIA L4 24 GB (AWS G6, Azure NVadsA10/NCads equivalents, GCP G2)
VRAM
24 GB
System RAM
64 GB
Storage
1000 GB
CPU
8–16 vCPU
Form factor
Cloud instance
Indicative costUS$1 – US$2
indicative on-demand rental, USD per hour, Aug 2026 — verify against the provider price list for your region
The private-cloud counterpart of `onprem-small-24gb`: same model sizes, no capital outlay, and a region you choose explicitly. Running it continuously for a year usually costs more than buying the equivalent box, so it suits pilots, bursts and firms without a server room. The cloud provider becomes a data processor — a DPA and a documented region are required.
Apple Silicon workstation (unified memory)
GPU
Apple M4 Max (up to 128 GB unified) or M3 Ultra (up to 256 GB unified)
VRAM
96 GB
System RAM
128 GB
Storage
2000 GB
CPU
Apple M4 Max / M3 Ultra
Form factor
Workstation
Indicative costUS$4,500 – US$12,000
indicative build cost for the complete machine, USD, Aug 2026 — verify with a local supplier
Quiet, standard-socket power draw, and a large memory pool — about 75% of installed unified memory is addressable by the GPU, which is where the 96 GB figure comes from on a 128 GB M4 Max. Memory bandwidth, not capacity, sets the ceiling: excellent for a pilot, a partner workstation or a single heavy user, weak for 40 people at once. Runs Ollama, llama.cpp and MLX; does not run CUDA builds of vLLM.

Indicative costUSD · one-off plus monthly

Software licences: All components are open source and self-installed. Check each licence for use limits.; US$0
Model usage: Open weights served locally: no per-token charge. The GPU cost sits under infrastructure.; US$0
GPU server (24 GB class): Hardware profile onprem-small-24gb — indicative build cost, Aug 2026, verify locally.; US$4,000 – US$9,000
Implementation (6–15 FDE-days): 6–15 FDE-days at US$760–1940 per day, converted from the HK$6,000–15,000 band at the HKMA Linked Exchange Rate band of HK$7.75–7.85 to one US dollar. One-off; excludes internal staff time.; US$4,560 – US$29,100

10–50 people recording meetings, one CRM, one language plus occasional others.
Excludes legal advice on recording consent, which is a prerequisite rather than an option.
Every figure is an assessment from the inputs listed, not a quotation.

05Difficulty

3 / 5

A project week with Linux, Docker and GPU experience

06Skills

API integrationapi-integration: development
Compliance and governancecompliance-governance: compliance
Dockerdocker: infrastructure
Linuxlinux: operations
LLM inferencellm-inference: ml
Pythonpython: development
Speech to textspeech-to-text: ml
Workflow automationworkflow-automation: operations

07Deployment steps

7 steps

Commands are copied from each project’s own documentation, and the page they came from is linked under the step. 0 of 7 steps currently open an evidence record. The rest are linked to their source; §10 says which of those documents were fetched and which were fetched without their anchor being found — two different states, named differently there.

01
Settle recording consent and retention before building
Assessment
Who may be recorded, how they are told, how consent is captured and how long recordings and transcripts are kept. This is a legal question with a different answer in each jurisdiction you operate in, and it determines what the pipeline is allowed to do. Get it answered in writing first; everything below is cheaper than rebuilding around it later.
02
Transcribe with faster-whisper on the GPUversion-sensitive
Assessment
Install faster-whisper and run the large-v3 model with float16 on GPU (or int8 to save memory). Turn on the VAD filter to drop silence, and use batched transcription for throughput on a backlog of recordings.
```
pip install faster-whisper
```
```
from faster_whisper import WhisperModel

model_size = "large-v3"

# Run on GPU with FP16
model = WhisperModel(model_size, device="cuda", compute_type="float16")
```
from the faster-whisper README
Source documentation
03
Add speaker labels where the summary needs themversion-sensitive
Assessment
WhisperX adds alignment and diarisation on top of Whisper. "The easiest way to install WhisperX is through PyPi." Diarisation needs a separate model and, for some pipelines, accepting the model licence — check that before assuming it will run offline.
```
pip install whisperx
```
Source documentation
04
Summarise against a schemaversion-sensitive
Assessment
Decide the fields the CRM actually needs — summary, decisions, action items with owners and dates, next step, sentiment if it is genuinely used — and generate them under constrained decoding so every note has the same shape. Keep the transcript so any claim in the summary can be traced back to what was said.
```
docker run --runtime nvidia --gpus all \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=$HF_TOKEN" \
    -p 8000:8000 \
    --ipc=host \
    vllm/vllm-openai:latest \
    --model Qwen/Qwen3-0.6B
```
from the vLLM docs — substitute the summarisation model
Source documentation
05
Match the recording to the right CRM recordversion-sensitive
Assessment
Matching is where these projects fail quietly. Use the calendar invitee list or the dialled number rather than name matching on the transcript, and send anything ambiguous to a human queue. A note filed against the wrong client is worse than no note.
```
docker volume create n8n_data

docker run -it --rm \
 --name n8n \
 -p 5678:5678 \
 -e GENERIC_TIMEZONE="<YOUR_TIMEZONE>" \
 -e TZ="<YOUR_TIMEZONE>" \
 -e N8N_ENFORCE_SETTINGS_FILE_PERMISSIONS=true \
 -e N8N_RUNNERS_ENABLED=true \
 -v n8n_data:/home/node/.n8n \
 docker.n8n.io/n8nio/n8n
```
from the n8n docs
Source documentation
06
Require approval before anything reaches the CRM
Assessment
The person who was in the meeting reviews the note and presses save. It takes thirty seconds, it catches the misattributed action item, and it is the difference between a useful tool and a CRM full of plausible fiction.
07
Automate deletion
Assessment
Recordings, transcripts and summaries each need a retention period and an actual deletion job — usually the recording goes first and the approved note stays. Store the timestamps so the deletion can be proved, not just asserted.
```
CREATE EXTENSION vector;
```
only if you also want semantic search over transcripts
Source documentation

08Compliance considerations

Structured issue-spotting to support your own review — not legal advice. Verify against the cited primary sources and your counsel.

Applies everywhere

Consent · Personal data · Transparencyhigh
Recording a conversation is processing personal data about everyone in it, including the people who are not your employees. Consent or another lawful basis, and clear notice, are prerequisites — not a later refinement. Requirements differ materially by jurisdiction.
Retentionhigh
Three artefacts, three lifetimes: recording, transcript, approved note. Keeping recordings indefinitely because storage is cheap is the most common failure here.
Confidentiality · Professional secrecyhigh
Transcripts of client meetings inherit the confidentiality of the meeting. Restrict access to the people who were entitled to be in the room.
Human oversight · Bias and fairnessmedium
Transcription accuracy varies by accent and language, so a summary of a heavily accented speaker can be systematically worse. Measure per language before rollout, and require human approval regardless.

09Alternatives

A meeting-notes SaaS product
Joins the call, transcribes, summarises and writes to the CRM with no infrastructure at all. The right answer when the conversations are not sensitive.
- — Working within a day
- — Every recorded conversation goes to a processor — consent, retention and residency questions
- — Per-seat pricing that grows with the team
fireflies otter
Transcription only, no CRM integration
Transcribe and summarise into a shared folder, and let people paste what matters. Removes the matching problem and most of the build.
- — A fraction of the work
- — Nothing is filed automatically; adoption depends on habit
faster-whisper whisperx

No alternative recipe is published yet.

10Evidence

0 of 6 fetched

Not yet in the ledger

WhisperX — README (setup and diarisation)
Readme
not yet fetched
faster-whisper — README (installation and usage)
Readme
not yet fetched
vLLM — Using Docker
Docs
not yet fetched
vLLM — Structured Outputs
Docs
not yet fetched
n8n — Install with Docker
Docs
not yet fetched
pgvector — README (installation and getting started)
Readme
not yet fetched

11Community

Deployed this stack, or hit something this page does not cover? Corrections, sources and implementation reports are what keep a recipe worth reading.

Improve this page

From the field

0 deployments · 0 questions

Nobody has reported deploying this here yet, and no question has been opened against this page. Both appear once a reviewer accepts them.

12Hire an FDE

If you would rather not build it, we can introduce a forward-deployed engineer who has deployed this stack before. The enquiry form starts from this recipe.

Hire an FDE for this stack

01Objective

Reference deployments using this recipe

02Recommended stack

Architecture and data flow

03Suitable for

04Hardware

On-premise single 24 GB GPU server

Cloud GPU instance — 1 × NVIDIA L4 24 GB

Apple Silicon workstation (unified memory)

Indicative cost·USD · one-off plus monthly

05Difficulty

06Skills

07Deployment steps

Settle recording consent and retention before building

Transcribe with faster-whisper on the GPUversion-sensitive

Add speaker labels where the summary needs themversion-sensitive

Summarise against a schemaversion-sensitive

Match the recording to the right CRM recordversion-sensitive

Require approval before anything reaches the CRM

Automate deletion