A model is still running

A model is still running or downloading. Leaving will stop the current operation. Stay on this page to let it finish.

Skip to content
LocalAI.toolsBeta

Join the waitlist

Get an email when the new apps and model runs launch.

No account needed.

Whisper 3

Run speech recognition in your browser. Unlimited runs.

Browser inferencePowered by WebAI · Private · Unlimited runs
About this model

Transcribe multilingual speech, translate speech into English text, or add optional segment or word timestamps using Whisper Large V3 Turbo. Returns text and timing data; it does not label speakers. Supports optional word-level timestamps.

Created by
OpenAI
License
MIT
Model released
Unknown
Source repository created
Oct 1, 2024
Model weights · Q4F16
563.6 MB

Best for

  • Private, client-side transcription of short and medium-length speech recordings.
  • Generating word- or segment-timestamped transcripts for captions and search.
  • Translating supported spoken languages into English text.
  • Prototyping browser speech-recognition features without a hosted inference API.

Keep in mind

  • Large V3 Turbo requires very large downloads, substantial memory, and capable hardware despite its reduced decoder depth.
  • The reduced decoder depth can slightly lower transcription quality compared with the full Whisper Large V3 model.
  • Word timestamps are approximate DTW alignments and may drift around silence, noise, or overlapping speech.
  • Accuracy varies by language, accent, recording quality, background noise, and domain vocabulary.
  • Long recordings are chunked and may contain boundary repetition or omissions.
  • Generated transcripts must not be treated as authoritative for safety-critical decisions.

Processed on your deviceInference runs in this browser. Input/output sharing requires consent for each run. Basic run logs include device information and approximate connection location. Model and runtime files download from external hosts. Saved downloads may be cleared by your browser.

Sources & technical details
Original model
openai/whisper-large-v3-turbo
Release verification
Unknown
Release evidence
Exact original checkpoint public release date has not been verified. Repository creation, training, paper, artifact conversion and update dates are not release evidence.
Source updated
Oct 4, 2024
Browser model files
onnx-community/whisper-large-v3-turbo_timestamped
Files updated
Mar 6, 2025
Artifact revision
b3f77bf9a8c4d5ea3415827033d1ffea7955fd9a

Input

Audio ≤500 MB · Video ≤2 GB

Sample result

Whisper Large V3 Turbo — Timestamped

Settings

Spoken-language code from the pinned Transformers.js Whisper tokenizer. Choose the actual spoken language; quality varies by language. The worker defaults to English when omitted. Choose the language actually spoken; a mismatched language can harm transcription. Automatic detection is available only when listed or explicitly allowed. This setting does not select a translation target.
Preserve the source language or translate speech to English. Choose a task matching the desired output. Caption tasks produce descriptions, OCR reads text, detection/grounding locates regions when offered. For speech, transcribe preserves the spoken language and translate outputs English; unsupported task values are not available.
Word timestamps, segment timestamps, or text only. word returns word-level timing, true returns segment timing, and false returns text only. Finer timing adds output and alignment work; timestamps are estimates.
Advanced settings 5
Use prior decoded tokens as context across chunks. On can improve continuity but may propagate earlier errors or repetitions. Off makes chunks more independent and can lose cross-chunk context.
Probability threshold used to classify a segment as silence. Lower values mark more segments as silence and may drop quiet speech. Higher values keep more uncertain segments and may transcribe background noise. It interacts with the decoder’s other silence checks.
Beam-search width; larger values trade speed for search quality. The listed range is a frontend selection policy, not a claimed model or runtime maximum; the adapter may accept values outside it. 1 uses a single search path. Higher values explore more alternatives and can improve wording, but use more time and memory and can still produce incorrect text.
Long-audio chunk duration in seconds. Shorter chunks reduce context and per-chunk memory but add boundaries. Longer chunks retain more context within the supported window and can use more memory.
Overlap on each side; cannot exceed half the chunk duration. Lower overlap reduces duplicated work but may reduce accuracy near boundaries. Higher overlap adds context and processing work; it must not exceed half the chunk duration.
Record a clip, then run inference. This does not send live audio to the model.

Output

Your transcript starts here

Choose audio or try a sample, then run transcription. Copy the text or export subtitles when ready.

Benchmark evidence

Saved worker evidence is separate from benchmarks run on your device. It does not guarantee speed or quality on every browser.

Whisper Large V3 Turbo — Timestamped

16 configurations tested
Passed9
Failed7
TestedSep 7, 2026
RuntimeTransformers.js 4.2.0

Quality criteria

No explicit quality gate recorded. Review per-case evidence; full certification is pending.

PrecisionDeviceResult
Q4F16GPU · WebGPUAcceleratedPassedFast
FP16GPU · WebGPUAcceleratedPassedFast
Q4GPU · WebGPUAcceleratedPassedFast
Q8GPU · WebGPUAcceleratedPassedVery slow
UINT8GPU · WebGPUAcceleratedPassedVery slow
INT8GPU · WebGPUAcceleratedPassedVery slow
BNB4GPU · WebGPUAcceleratedPassedVery slow
BNB4CPU · WASMPassedVery slow
Q4CPU · WASMPassedVery slow
Q4F16CPU · WASMFailed—
Q8CPU · WASMFailed—
INT8CPU · WASMFailed—
UINT8CPU · WASMFailed—
FP16CPU · WASMFailed—
FP32CPU · WASMFailed—
FP32GPU · WebGPUFailed—
Test setup & sample details
Quality gate
No explicit quality gate recorded. Review the per-case quality evidence; full certification is pending.
Tested
2026-09-07T19:18:13.382Z
Environment
Headless Chrome 151 on macOS; Transformers.js 4.2.0
Runtime
Transformers.js 4.2.0
Artifact revision
b3f77bf9a8c4d5ea3415827033d1ffea7955fd9a
Fixture
{
  "file": "whisper-standard.mp3",
  "url": "https://f005.backblazeb2.com/file/hubters-webai-public/assets/benchmarks/whisper-standard.mp3",
  "size": 471925,
  "sha1": "5faad1507b5d6bc61a39ca3184db5a7b554a7479",
  "durationSeconds": 29.49225
}
Rerun when
model artifact revision change; Transformers.js or ONNX Runtime version change; benchmark fixture or quality-gate change; target browser/backend behavior change
Per-case quality & failure details
[
  {
    "precision": "q4",
    "device": "wasm",
    "status": "pass",
    "quality": {
      "grade": "excellent",
      "wordErrorRate": 0.1364,
      "gibberish": false
    },
    "runtime": {
      "loadMs": 78979.7,
      "inferenceMs": 131504.2,
      "totalMs": 210483.9,
      "grade": "very-slow",
      "realTimeFactor": 4.4589,
      "loadGrade": "slow"
    }
  },
  {
    "precision": "q4",
    "device": "webgpu",
    "status": "pass",
    "quality": {
      "grade": "excellent",
      "wordErrorRate": 0.1364,
      "gibberish": false
    },
    "runtime": {
      "loadMs": 78862.2,
      "inferenceMs": 4600.6,
      "totalMs": 83462.9,
      "grade": "excellent",
      "realTimeFactor": 0.156,
      "loadGrade": "slow"
    }
  },
  {
    "precision": "bnb4",
    "device": "wasm",
    "status": "pass",
    "quality": {
      "grade": "excellent",
      "wordErrorRate": 0.1364,
      "gibberish": false
    },
    "runtime": {
      "loadMs": 75953.2,
      "inferenceMs": 117840.2,
      "totalMs": 193793.5,
      "grade": "very-slow",
      "realTimeFactor": 3.9956,
      "loadGrade": "slow"
    }
  },
  {
    "precision": "bnb4",
    "device": "webgpu",
    "status": "pass",
    "quality": {
      "grade": "excellent",
      "wordErrorRate": 0.1364,
      "gibberish": false
    },
    "runtime": {
      "loadMs": 77303.5,
      "inferenceMs": 107462.5,
      "totalMs": 184766,
      "grade": "very-slow",
      "realTimeFactor": 3.6438,
      "loadGrade": "slow"
    }
  },
  {
    "precision": "q4f16",
    "device": "wasm",
    "status": "fail",
    "quality": "Not embedded",
    "runtime": {
      "loadMs": null,
      "inferenceMs": null,
      "totalMs": 156162.1,
      "grade": "failed"
    },
    "failure": "onnx-memory-error"
  },
  {
    "precision": "q4f16",
    "device": "webgpu",
    "status": "pass",
    "quality": {
      "grade": "excellent",
      "wordErrorRate": 0.1364,
      "gibberish": false
    },
    "runtime": {
      "loadMs": 153483.4,
      "inferenceMs": 3570.7,
      "totalMs": 157054.1,
      "grade": "excellent",
      "realTimeFactor": 0.1211,
      "loadGrade": "slow"
    }
  },
  {
    "precision": "q8",
    "device": "wasm",
    "status": "fail",
    "quality": "Not embedded",
    "runtime": {
      "loadMs": null,
      "inferenceMs": null,
      "totalMs": 91643.1,
      "grade": "failed"
    },
    "failure": "onnx-quantization-error"
  },
  {
    "precision": "q8",
    "device": "webgpu",
    "status": "pass",
    "quality": {
      "grade": "excellent",
      "wordErrorRate": 0.1364,
      "gibberish": false
    },
    "runtime": {
      "loadMs": 112050.8,
      "inferenceMs": 92618.4,
      "totalMs": 204669.2,
      "grade": "very-slow",
      "realTimeFactor": 3.1404,
      "loadGrade": "slow"
    }
  },
  {
    "precision": "int8",
    "device": "wasm",
    "status": "fail",
    "quality": "Not embedded",
    "runtime": {
      "loadMs": null,
      "inferenceMs": null,
      "totalMs": 95212.9,
      "grade": "failed"
    },
    "failure": "onnx-quantization-error"
  },
  {
    "precision": "int8",
    "device": "webgpu",
    "status": "pass",
    "quality": {
      "grade": "excellent",
      "wordErrorRate": 0.1364,
      "gibberish": false
    },
    "runtime": {
      "loadMs": 110754.5,
      "inferenceMs": 94937.9,
      "totalMs": 205692.4,
      "grade": "very-slow",
      "realTimeFactor": 3.2191,
      "loadGrade": "slow"
    }
  },
  {
    "precision": "uint8",
    "device": "wasm",
    "status": "fail",
    "quality": "Not embedded",
    "runtime": {
      "loadMs": null,
      "inferenceMs": null,
      "totalMs": 96533,
      "grade": "failed"
    },
    "failure": "onnx-quantization-error"
  },
  {
    "precision": "uint8",
    "device": "webgpu",
    "status": "pass",
    "quality": {
      "grade": "excellent",
      "wordErrorRate": 0.1364,
      "gibberish": false
    },
    "runtime": {
      "loadMs": 111100.9,
      "inferenceMs": 93238.7,
      "totalMs": 204339.6,
      "grade": "very-slow",
      "realTimeFactor": 3.1615,
      "loadGrade": "slow"
    }
  },
  {
    "precision": "fp16",
    "device": "wasm",
    "status": "fail",
    "quality": "Not embedded",
    "runtime": {
      "loadMs": null,
      "inferenceMs": null,
      "totalMs": 176594.5,
      "grade": "failed"
    },
    "failure": "onnx-memory-error"
  },
  {
    "precision": "fp16",
    "device": "webgpu",
    "status": "pass",
    "quality": {
      "grade": "acceptable",
      "wordErrorRate": 0.1515,
      "gibberish": false
    },
    "runtime": {
      "loadMs": 174688,
      "inferenceMs": 4018.3,
      "totalMs": 178706.4,
      "grade": "excellent",
      "realTimeFactor": 0.1362,
      "loadGrade": "slow"
    }
  },
  {
    "precision": "fp32",
    "device": "wasm",
    "status": "fail",
    "quality": "Not embedded",
    "runtime": {
      "loadMs": null,
      "inferenceMs": null,
      "totalMs": null,
      "grade": "failed"
    },
    "failure": "timeout"
  },
  {
    "precision": "fp32",
    "device": "webgpu",
    "status": "fail",
    "quality": "Not embedded",
    "runtime": {
      "loadMs": null,
      "inferenceMs": null,
      "totalMs": 119554.8,
      "grade": "failed"
    },
    "failure": "artifact-fetch-error"
  }
]

Something not quite right?

Tell us what happened with Whisper 3. Loading, results, or the page itself—we’re here for all of it.

Attachments Optional

Up to 3 files · 50 MB each

Browser, OS, CPU/GPU, memory, screen, storage, and network details are included automatically. Model inputs, results, and browsing history are not collected.