Best free AI transcription: 98 local model tests

Which route fits your recording?

• Files already on your computer: local Whisper avoids sending the audio to a transcription API. It needs installation and model downloads; this test used Python 3.12 on Windows, CPU int8 and two threads.
• Meetings and shared notes: recording permission, speaker labels, collaboration, retention and integrations are separate requirements. See the meeting-notetaker guide; no hosted tool was measured in this experiment.
• Accuracy-sensitive text: retain the audio and check names, numbers, technical terms and unexpected phrases. Neither model's output is a verified transcript.

There is no metered transcription API charge for these local runs, but hardware, power and setup still have a cost. The runtime documentation and model cards describe local use. This is not a review of every free plan.

The measured difference with added noise

Base.en had fewer counted edits in the noise-added condition on this subset, but not on every individual clip. The clean totals tied despite different errors. The per-clip answers below preserve those differences.

Recordings and first answers

96 speech outputs across 24 recordings.

Silence is not always an empty transcript

Both models returned You on one five-second digital-silence file with VAD off. That is one probe per model, not a measured hallucination rate. It is excluded from speech WER.

Another visible mismatch: base.en added Thank you very much after the corpus reference for original clip 237-126133-0001. The four insertions remain in the reported score; we have not independently re-adjudicated the audio.

Run the disclosed inputs locally

Extract the kit into an empty folder. Its Python runner checks every audio hash and pins the same model revisions and package versions. It does not upload recordings. Model downloads come from Hugging Face; no account is needed for these public files.

1. Create an isolated Python environmentExtract the kit into an empty folder and create a Python 3.12 virtual environment.
2. Install the pinned packagesInstall the exact package versions listed in requirements.txt.
3. Verify every disclosed inputRun reproduce.py with --verify-only to check all 49 audio hashes without inference.
4. Write a new result fileChoose tiny.en or base.en and a new output filename. The runner refuses to overwrite an existing result.

From that folder, Windows PowerShell with Python 3.12:

py -3.12 -m venv .venv
.venv\Scripts\python -m pip install -r requirements.txt
.venv\Scripts\python reproduce.py --verify-only
.venv\Scripts\python reproduce.py --model tiny.en --output my-tiny.jsonl
.venv\Scripts\python reproduce.py --model base.en --output my-base.jsonl

On macOS/Linux, use python3 -m venv .venv and .venv/bin/python for the remaining commands. That operating-system path has not been execution-tested here. Audio verification runs without the inference dependencies; transcription needs the installed packages and models.

The portable wrapper's input validation and inference interface were tested separately from the original measurement runner. It has not been used to rerun this entire benchmark or tested in a freshly installed environment. Matching settings do not guarantee byte-identical output on different hardware. Existing output files are never overwritten.

Method, provenance and boundaries

• Selection: the first 12 numeric speaker IDs and first two lexicographic utterance IDs per speaker from LibriSpeech clean/test. This convenience subset is 24 of 2,620 clips, not representative of all accents, devices or everyday speech.
• Frozen before inference: selection, model revisions, audio hashes and settings were locked at 21:14:45 UTC on 28 August 2026; the 98 first calls completed at 21:16:42 UTC. No output-based reselection or prompt tuning.
• Runtime: faster-whisper 1.2.1, CTranslate2 4.7.1, CPU int8, two threads, one worker, beam 5, temperature 0, English, VAD off and previous-text conditioning off. No prompt or vocabulary hints.
• Hardware and timing: Ryzen 7 8845HS; tiny then base, no warm-up or repeated timing trials, background load uncontrolled. Timings include file decoding and lazy segment consumption, but exclude model loading and download.
• Scoring: Jiwer 4.0.0; lowercase, remove punctuation, collapse spaces. WER = (S + D + I) / reference words. Totals sum clip counts rather than averaging percentages. Empty-reference silence is separate; WER can exceed 100%.
• Noise: seeded NumPy white noise at whole-clip 10 dB RMS SNR. Original FLAC bytes are unchanged; noise variants are float32 WAV, uniformly peak-scaled when necessary.
• Limits: no human listening/re-adjudication, confidence intervals, statistical superiority, diarization, overlapping speech, meetings or non-English test. Public-corpus overlap with training is unknown.

Some counted errors are spelling differences

In 260-123286-0000, both models write 15th where the corpus writes FIFTEENTH: one substitution. In 61-70968-0001, tiny.en writes not-so; punctuation removal joins it to notso, unlike reference NOT SO. US/UK spelling and compound words can also affect counts. The primary scores retain the frozen normalization.

Full public method | Input hashes, settings and model revisions | Original first outputs JSONL | Summary CSV.

Questions about free transcription

Sources and reuse

Audio and reference text: LibriSpeech, Panayotov, Chen, Povey and Khudanpur, via the pinned OpenSLR dataset, CC BY 4.0. Noise-added files are ToolGlance modifications; source audio is not our own recording.

Models: tiny.en and base.en. Runtime: faster-whisper 1.2.1. Scorer: Jiwer. Attribution and reuse terms.

Cite ToolGlance's 28 August 2026 local transcription test for the disclosed subset, settings and observed edit counts. Keep the sample and normalization limits beside reused figures; do not cite it as a winner across services, accents or meeting workflows.

Other directory tools and related reading

The linked hosted tools were not tested here; their directory listings are not measurements from this experiment.

• Otter.ai
• Descript
• Fathom
• Fireflies.ai
• Avoma
• Sembly AI
• Fellow
• tl;dv
• Best free AI meeting notetaker in 2026
• Best AI to remove background noise from audio and calls
• Best AI Meeting Notetaker in 2026: Fathom vs Otter vs Fireflies vs tl;dv
• State of AI Video Generation 2026
• State of AI Image Generation 2026
• State of AI in Marketing 2026