Agentic WeeklyGet the weekly
SYSTRANv1.2.1Tested (smoke test)

faster-whisper: what’s new in v1.2.1 and our verdict

Speech-to-text library that runs OpenAI Whisper models with CTranslate2.

Official site

Release Ruler for faster-whisper: marked at its Tested release date on the timeline of every AI model we track.

At a glance

Evidence
Tested (smoke test)
Released
Oct 31, 2025
Last verified
Oct 25, 2026
Pricing
Partly verified
Best for
AI voice and speech

Evidence and ranking

Evidence behind this verdict

  1. Sources read1 cited source, latest fetched Oct 25, 2026. (done)
  2. Facts re-checkedLast verified Oct 25, 2026. (done)
  3. Run by usSmoke test on Oct 25, 2026 (scope below). (done)

Passed Test log result: Passed. Read the log

Where we rank it (editorial)

Listed here but not yet ranked. Ranks are editorial picks, not benchmark scores.

Pricing: Open source; no pricing page read. (checked 2026-10-25)

Our verdict

Fast local transcription; on CPU use float32 and check int8 against a reference before trusting it.

Good for

  • Local, private transcription
  • Developers building voice pipelines on CPU

Skip if

  • You need diarisation or a managed service

Tested. Tested 2026-10-25 (America/Chicago). faster-whisper 1.2.1 transcribed 41 s of speech at 2.4 percent word error with the base.en model in float32 (deterministic), but the int8 setting gave 2 to 60 percent on identical audio across runs. Test environment: Linux x86_64 virtual machine (Debian 13), 8 vCPU, 15 GiB RAM, no GPU; no vendor account, login or API key. No language model was involved in this test, so it says nothing about any model, vendor-hosted or local. Not tested: Other audio: one synthetic (Piper) voice, English only, no noise, no accents, no long recordings; GPU, other compute types (float16, int8_float16) and the larger models; Not a benchmark; this is a smoke test on 41 seconds of audio. These are smoke tests, not a benchmark. Full log: /tests/faster-whisper.

Evidence
Tested (smoke test, 2026-10-25, no language model involved). A smoke test is not a benchmark.
What we ran
faster-whisper 1.2.1 transcribed 41 s of speech at 2.4 percent word error with the base.en model in float32 (deterministic), but the int8 setting gave 2 to 60 percent on identical audio across runs.
Model used
None; no language model was involved
Environment
Linux x86_64 virtual machine (Debian 13), 8 vCPU, 15 GiB RAM, no GPU; no vendor account, login or API key
Not tested
Other audio: one synthetic (Piper) voice, English only, no noise, no accents, no long recordings; GPU, other compute types (float16, int8_float16) and the larger models; Not a benchmark; this is a smoke test on 41 seconds of audio.
Date
(America/Chicago)

Best for

Key facts

faster-whisper reimplements Whisper inference with CTranslate2 for faster CPU and GPU transcription. It is MIT-licensed; v1.2.1 was tagged on 2025-10-31.

  • License: MIT (per GitHub)
  • Platforms: Python; CPU and NVIDIA GPU
  • Pricing model: Open source
  • Free tier: Open source, free to use
  • Paid plans: Hosted or enterprise offerings not read
  • GitHub: SYSTRAN/faster-whisper: 26,138 stars on 2026-10-25, last push 2026-10-25
  • Latest release we saw: v1.2.1, 2025-10-31
  • How to start: uv pip install faster-whisper av==18.1.0.
  • v1.2.1 (2025-10-31): the newest release on the repository when we read it on 2026-10-25.

How to set it up: Get started with faster-whisper

Install steps are exactly what we ran on 2026-10-25; the test log lists what was and was not covered.

  1. Install and pin av

    In our run an unpinned av release broke first use; pinning fixed it.

    uv pip install faster-whisper av==18.1.0

  2. Transcribe

    Use float32 on CPU if you need repeatable output.

    WhisperModel("base.en", device="cpu", compute_type="float32")

Pros and cons

Pros

  • MIT license
  • Repeatable word error rate in float32 on our test audio

Cons

  • int8 on CPU gave 2.4 to 59.7 percent word error on the same audio run to run in our test
  • An unpinned dependency (av 19.0.2) broke first use

Alternatives and comparisons

  • Piper: Fast local neural text-to-speech that runs on CPU.
  • Ollama: The common way to run open models locally on macOS, Windows and Linux, with a cloud option and an MIT-licensed open-source core.
  • llama.cpp: LLM inference in C/C++: the open-source engine many local-model tools build on; MIT-licensed with near-daily build tags.
  • AssemblyAI: Speech AI APIs for transcription and understanding; start free, pay as you go.

FAQ

What is faster-whisper?

A library that runs Whisper speech-to-text models with CTranslate2 for faster inference.

Is int8 safe on CPU?

Check it first: in our test the same audio gave between 2.4 and 59.7 percent word error across int8 runs, while float32 was stable. See /tests/faster-whisper.

Why did faster-whisper fail on first use?

An unpinned av 19.0.2 raised a TypeError in our run; av==18.1.0 fixed it.

What we have not verified

  • GPU speed
  • Non-English audio, noisy audio, long recordings

Sources

Last verified · Agentic Weekly

  • AssemblyAI: Speech AI APIs for transcription and understanding; start free, pay as you go.
  • Cartesia: Realtime voice AI: text-to-speech, speech-to-text in 44 languages and voice agents, with a free plan and credit tiers.
  • Deepgram: Voice AI APIs for speech-to-text, text-to-speech and agents with $200 in free credit.

How to cite this page

Agentic Weekly, "faster-whisper v1.2.1: what's new, verdict & setup", last updated 2026-10-25, https://theagenticweekly.com/tools/faster-whisper

Keep the Reported or Tested label when you quote a verdict. Plain-text version: https://theagenticweekly.com/tools/faster-whisper.md. More for AI assistants: /ai, /tools.json, /llms.txt.