A plain transcript is a wall of words. A transcript with speaker diarization tells you who said what — which is the difference between usable meeting notes and a confusing blob of text. Here are the best free ways to get speaker-labelled AI transcription in 2026, from a single command-line pipeline to hosted APIs with generous free tiers.
What speaker diarization actually does
Speaker diarization answers one question: "who spoke when?" It segments audio into speech regions, builds a voiceprint embedding for each segment, and clusters similar voices into speaker labels like SPEAKER_00 and SPEAKER_01. It does not know the speakers' real names — it just distinguishes voice A from voice B consistently through the file.
Accuracy is measured with Diarization Error Rate (DER): missed speech, false alarms, and speaker confusion. Open-source pyannote.audio 3.1 — the engine inside most free tools — scores 11–19% DER on standard benchmarks, competitive with commercial APIs. That is the engine to look for.
1. WhisperX — the best all-in-one free pipeline
WhisperX wraps OpenAI's Whisper (transcription) around pyannote.audio (diarization) and adds wav2vec2 forced alignment for word-level timestamps — all in one command. Output is a transcript where every segment carries a speaker label and word-accurate timings, exportable as JSON, SRT, TXT, TSV, or VTT.
whisperx meeting.mp3 --model large-v2 --diarize --language en --hf_token YOUR_TOKENThe catch: you need a GPU (a T4 or better) and a free HuggingFace token to download the pyannote diarization model. The popular workaround is Google Colab's free GPU tier — run the whole pipeline in a browser notebook with no local installs. For batch processing of interviews and podcasts on your own hardware, nothing free beats it.
2. pyannote.audio + faster-whisper — the DIY research stack
If WhisperX feels like a black box, build the two stages yourself: faster-whisper for transcription, pyannote.audio's community pipeline for diarization. Same 11–19% DER, same free and open models, but you control chunking, the number of speakers, and how the two streams merge.
This is the stack for developers and researchers: long-form files (a 4-hour meeting is routine), on-prem or air-gapped deployments, and privacy-critical audio that must never leave your machine. Expect slower processing on CPU and some Python/PyTorch fluency. If you redistribute output produced with the default pyannote community diarization pipeline, note its CC-BY-4.0 attribution requirement.
3. AssemblyAI — generous free tier on a hosted API
If you would rather call an API than manage GPUs, AssemblyAI's free tier covers 185 hours per year of pre-recorded audio, and speaker diarization is a standard parameter on the transcript endpoint. No infrastructure, no model downloads, no token management — upload audio, get back speaker-labelled text with timestamps in under a minute per hour of audio.
The trade-off is privacy and the annual cap: audio goes to their servers, and heavy users will hit the ceiling. For solo creators, podcasters, and small teams transcribing a few interviews a month, it is effectively free forever.
4. Deepgram — fast hosted diarization with a free tier
Deepgram's speech API includes speaker diarization as a built-in feature with low-latency turnaround — tens of seconds for an hour of audio. The free tier is enough to evaluate thoroughly and handle light ongoing use. It is the pick when speed matters: live-adjacent workflows, rapid interview turnaround, or feeding speaker-labelled transcripts into downstream apps.
Choose between AssemblyAI and Deepgram on turnaround time and pricing once you exceed free tiers; on raw diarization quality they sit in the same band (7–9% DER territory on clean audio per 2026 comparisons).
5. Google Recorder — free, on-device, and already in your pocket
Pixel's Recorder app transcribes speech on-device in real time and automatically labels distinct speakers, with no subscription and no cloud upload. It is limited by platform (Pixel phones) and by its consumer-app framing — you record meetings live rather than uploading arbitrary files — but for quick interviews, lectures, and conversations, it is the zero-setup answer.
Apple has been moving the same way: macOS 26's native SpeechAnalyzer framework exposes speaker change detection to apps, so free macOS transcription utilities are starting to ship on-device speaker labelling with zero new dependencies and zero network calls. If you live in the Apple ecosystem, watch this space — the free, private option is arriving natively.
6. NVIDIA NeMo — the open-source production engine
NVIDIA's NeMo toolkit is open source and includes state-of-the-art diarization modules that run fastest on NVIDIA GPUs. It is the pick when WhisperX's convenience matters less than throughput: large backlogs of calls, multi-hour conference archives, or pipelines where you already have NVIDIA infrastructure.
NeMo is more toolkit than turnkey tool — expect to write Python, configure manifests, and tune the pipeline. For teams with GPUs and scale, it is free and enterprise-grade.
7. Speakr — the self-hosted transcription platform
Speakr is an open-source, Docker-based transcription platform that runs entirely on your own hardware. It combines system-audio capture, WhisperX-powered transcription with speaker diarization, and even podcast-style summaries via local LLMs — no per-minute fees, no cloud retention policies.
Recent versions add multi-platform system audio capture and webhooks for automation tools like Zapier or n8n, so you can build a pipeline that records a Zoom call, transcribes it locally with speaker labels, and drops the summary into your project tracker. Ideal if you want Otter.ai-style convenience without Otter.ai-style subscriptions.
8. Vosk + custom diarization — the CPU-only edge option
For CPU-only machines and edge devices, Vosk provides fast offline speech recognition with no GPU and no cloud. Diarization takes extra wiring (typically a pyannote or clustering stage on top), and DER runs higher — 18–30% on noisy multi-speaker audio — but for quiet two-speaker recordings on a laptop, it is a legitimate free route. Consider it the fallback when you have no GPU, no free Colab quota, and cannot send audio to the cloud.
How to pick the right one
Match the tool to your constraints, not to a leaderboard:
- Have a GPU or free Colab access? WhisperX gives you the best accuracy-per-dollar (zero dollars) for batch files.
- Building a product? AssemblyAI or Deepgram on the free tier, then paid per hour once you scale — APIs are 30–60% cheaper per hour than running your own GPUs only after serious volume.
- Privacy-first or air-gapped? WhisperX, pyannote.audio + faster-whisper, NeMo, or Speakr — nothing leaves your machine.
- Just need it done now on a phone? Google Recorder on Pixel, or macOS 26's native speaker detection on Apple hardware.
- No GPU at all? Vosk + diarization for offline CPU work, or a hosted free tier.
Quick start: WhisperX on a free Colab GPU
The fastest zero-cost path to a speaker-labelled transcript:
1. Create a free HuggingFace account and generate an access token. Accept the license for the pyannote speaker-diarization model on its HuggingFace page — this is required for the download.
2. Open a new Google Colab notebook and switch the runtime to a free T4 GPU.
3. Install WhisperX and its dependencies in the notebook, store your HuggingFace token in Colab's secrets.
4. Upload your audio and run: whisperx yourfile.mp3 --language en --diarize --hf_token $HF_TOKEN --model large-v2.
5. Download the JSON or SRT output — every segment is labelled SPEAKER_00, SPEAKER_01, and so on.
Expect roughly 3–6 minutes of processing per hour of audio on a T4. For files longer than 10 minutes, chunk them into segments first to reduce GPU memory pressure, which also improves diarization accuracy.