Elastic model: thewhisper-large-v3-turbo

The project GitHub: TheWhisper · TheStage Apple SDK

Original model: openai/whisper-large-v3-turbo (OpenAI)

TheWhisper-Large-V3-Turbo is a fine-tuned, high-performance variant of OpenAI’s Whisper Large V3 model — optimized by TheStage AI for real-time, low-latency, and low-power speech-to-text (ASR) inference across multiple platforms, including NVIDIA GPUs and Apple Silicon (CoreML / Neural Engine).

Overview

---

ElasticModels are the models produced by TheStage AI ANNA: Automated Neural Networks Accelerator. ANNA allows you to control model size, latency and quality with a simple slider movement, routing different compression algorithms to different layers. For each model, we have produced a series of optimized models:

• XL: Mathematically equivalent neural network, optimized with our DNN compiler.
• L: Near lossless model, with less than 1% degradation obtained on corresponding benchmarks.
• M: Faster model, with accuracy degradation less than 1.5%.
• S: The fastest model, with accuracy degradation less than 2%.

How to run this model:

Models can be accessed via TheStage AI Python SDK (ElasticModels), the TheStage Apple SDK (Swift / Flutter), or deployed as Docker containers with REST API endpoints (see Deploy section).

System Requirements

---

NVIDIA

Apple Silicon (TheStage Apple SDK)

Simulator is not supported — run on real Apple Silicon hardware.

TheStage AI Access Token Setup

---

Install TheStage AI CLI and setup API token:

pip install thestage
thestage config set --access-token <YOUR_ACCESS_TOKEN>

For the Apple SDK, create a token at app.thestage.ai and pass it to TheStageAI.shared.initialize(apiToken:) (Swift) or TheStageFlutterSDK.initialize(api_token:) (Flutter). Token is checked online in initialize (once per app process when reachable). Inference runs fully on-device. Offline initialize fails — reconnect and call initialize again.

TheStage Apple SDK

---

On-device transcription for iOS and macOS. The SDK downloads CoreML engines from this Hugging Face repo, auto-selects Neural Engine / GPU / CPU, and exposes batch infer plus push-based live streaming. No server in the hot path.

Docs: TheStage Apple SDK · Whisper · Repo: TheStage Apple SDK

Installation (SwiftPM)

In Xcode: File → Add Package Dependencies…, paste https://github.com/TheStageAI/AppleSDK.git, and add the TheStageSDK product. Or in Package.swift:

.package(
    url: "https://github.com/TheStageAI/AppleSDK.git",
    exact: Version(1, 1, 0)
)

Swift — batch transcription

import TheStageSDK

let ai = TheStageAI.shared
try await ai.initialize(apiToken: "th_…")

let stt = try await WhisperPipeline(
    engines_path: "TheStageAI/thewhisper-large-v3-turbo"
)

// audio: 16 kHz mono Float, samples in [-1.0, 1.0]
let result = stt.infer(audio: audio_samples, language: "en")
print(result.text)

Long audio is split into the bundle's window size (10 s for the shipping turbo engines). Optional constructor knobs: device ("npu" / "gpu" / "cpu"), overlap_seconds, use_internal_vad.

Swift — live streaming

let streamer = stt.open_streamer(language: "en")

let captions = Task {
    for await text in streamer.partials {
        print("partial: \(text)")   // committed-so-far, grows monotonically
    }
}

for await frame in microphone_frames {   // [Float] @ 16 kHz mono
    streamer.send(frame)
    if vad_detected_pause { streamer.flush() }
}

let final_text = await streamer.finish()
await captions.value
print("final: \(final_text)")

partials is for live UI; finish() is the authoritative end-of-turn transcript. Call flush() at VAD pauses to keep latency flat on long turns; cancel() for barge-in.

Flutter (iOS)

# pubspec.yaml
dependencies:
  thestage_apple_sdk:
    git:
      url: https://github.com/TheStageAI/AppleSDK.git
      path: plugin/thestage_apple_sdk
      ref: 1.1.0

import 'package:thestage_apple_sdk/thestage_apple_sdk.dart';
import 'dart:typed_data';

await TheStageFlutterSDK.initialize(api_token: 'th_…');

await TheStageFlutterSDK.start_model(
  model_name: 'stt',
  engines_path: 'TheStageAI/thewhisper-large-v3-turbo',
);

// audio_samples: Float32List, 16 kHz mono, samples in [-1.0, 1.0]
final result = await TheStageFlutterSDK.infer(
  model_name: 'stt',
  input_json: {
    'audio': audio_samples,
    'language': 'en',
  },
);
print(result[0]['transcription']);

JSON response keys: transcription (String), token_count (Int), decode_seconds (Double), optional tokens ([Int]).

Audio contract (Apple)

Prefetch / progress

let engines_dir = try await ai.prefetch_engines(
    repo_id: "TheStageAI/thewhisper-large-v3-turbo"
)
let stt = try await WhisperPipeline(engines_path: engines_dir)

// Optional load progress:
let stt = try await WhisperPipeline(
    engines_path: "TheStageAI/thewhisper-large-v3-turbo",
    on_load_progress: { p in
        print("[\(p.model)] \(p.phase) \(Int(p.fraction * 100))%")
    }
)

Apple on-device latency

Release build on Apple M2 Max (NPU / ANE), macOS 26.2 (guidance, not an SLA):

Measured on a short ~2.6 s utterance; shipping windows are 10 s. Always bench release builds on device.

ElasticModels

---

Elastic Models provides the same interface as HuggingFace Transformers. Here is an example of how to use the thewhisper-large-v3-turbo model.

Installation

pip install 'thestage-elastic-models[nvidia]' \
    --extra-index-url https://thestage.jfrog.io/artifactory/api/pypi/pypi-thestage-ai-production/simple
pip install datasets==3.6.0 librosa soundfile  # only needed to load audio for the example below

Usage

import torch
from elastic_models.transformers import AutoModelForSpeechSeq2Seq
from transformers import AutoProcessor

model_name = "TheStageAI/thewhisper-large-v3-turbo"
hf_token = ''
device = torch.device("cuda")

processor = AutoProcessor.from_pretrained(
    model_name, token=hf_token
)
model = AutoModelForSpeechSeq2Seq.from_pretrained(
    model_name,
    token=hf_token,
    torch_dtype=torch.float16,
    mode='S'
).to(device)

# Load audio file
from datasets import load_dataset
dataset = load_dataset(
    "hf-internal-testing/librispeech_asr_dummy",
    "clean", split="validation"
)
audio_sample = dataset[0]["audio"]

# Process audio
input_features = processor(
    audio_sample["array"],
    sampling_rate=audio_sample["sampling_rate"],
    return_tensors="pt"
).input_features.to(device, dtype=torch.float16)

# Generate transcription
with torch.inference_mode():
    predicted_ids = model.generate(input_features)

transcription = processor.batch_decode(
    predicted_ids, skip_special_tokens=True
)[0]

print(f"Transcription: {transcription}")

TheWhisper SpeechKit

---

TheWhisper can also be used via thestage_speechkit for NVIDIA and Apple Silicon inference from Python, including real-time streaming.

Installation

# Install ffmpeg (required for audio processing)
# Ubuntu/Debian: apt install ffmpeg
# macOS: brew install ffmpeg
git clone https://github.com/TheStageAI/TheWhisper.git
cd TheWhisper
pip install .[nvidia]   # or pip install .[apple] for Apple Silicon

For TheStage AI optimized engines (NVIDIA only):

pip install thestage-elastic-models[nvidia] --extra-index-url https://thestage.jfrog.io/artifactory/api/pypi/pypi-thestage-ai-production/simple

NVIDIA Usage

from thestage_speechkit.nvidia import ASRPipeline

model = ASRPipeline(
    model='TheStageAI/thewhisper-large-v3-turbo',
    model_size='S',
    chunk_length_s=15,
    batch_size=32,
    device='cuda'
)

result = model(
    "path_to_your_audio.wav",
    chunk_length_s=15,
    generate_kwargs={'do_sample': False, 'num_beams': 1, 'use_cache': True}
)

print(result["text"])

Apple Silicon Usage (Python)

For shipping iOS / macOS apps, prefer the TheStage Apple SDK above. SpeechKit remains available for Python notebooks and macOS scripting.

from thestage_speechkit.apple import ASRPipeline

model = ASRPipeline(
    model='TheStageAI/thewhisper-large-v3-turbo',
    model_size='S',
    chunk_length_s=10
)

result = model(
    "path_to_your_audio.wav",
    chunk_length_s=10,
    generate_kwargs={'do_sample': False, 'num_beams': 1, 'use_cache': True}
)

print(result["text"])

Streaming

from thestage_speechkit.streaming import StreamingPipeline, MicStream, StdoutStream

streaming_pipe = StreamingPipeline(
    model='TheStageAI/thewhisper-large-v3-turbo',
    model_size='S',
    chunk_length_s=15,
    platform='apple',
    language='en'
)

mic_stream = MicStream(step_size_s=0.5)
output_stream = StdoutStream()

while True:
    chunk = mic_stream.next_chunk()
    if chunk is not None:
        approved_text, assumption = streaming_pipe(chunk)
        output_stream.write(approved_text, assumption)
    else:
        break

Quality Benchmarks

---

We have evaluated the models using the Hugging Face Open ASR Leaderboard methodology. For each model size (S, M, L, XL), we report Word Error Rate (WER) on standard English and multilingual speech recognition benchmarks.

Open ASR Leaderboard (English, WER %)

Multilingual (WER %)

Datasets

---

English (Open ASR Leaderboard)

• LibriSpeech Clean: Read English speech from audiobooks, recorded in clean conditions. Tests baseline transcription accuracy on clear, well-articulated speech.
• LibriSpeech Other: Read English speech from audiobooks with more challenging acoustic conditions, including noisier recordings and less common speakers.
• SPGISpeech: Financial earnings calls and presentations, featuring domain-specific terminology, spontaneous speech, and diverse speaker accents.
• TEDLium: TED conference talks covering a wide range of topics, with diverse speakers, presentation styles, and varying audio quality.
• VoxPopuli: European Parliament event recordings in multiple languages, featuring political discourse, formal speech, and multilingual speakers.
• GigaSpeech: Large-scale multi-domain English speech corpus from audiobooks, podcasts, and YouTube, representing diverse acoustic conditions and speaking styles.
• Earnings22: Corporate earnings calls with financial terminology, multiple speakers, and telephone-quality audio.
• AMI: Meeting recordings with overlapping speech, distant microphones, and natural conversational dynamics.

Multilingual

• CoVoST2: Common Voice Speech-To-Text 2. Built on Mozilla's Common Voice recordings, providing speech-to-text evaluation across 21 languages with diverse speakers, accents, and recording conditions.
• FLEURS: Few-shot Learning Evaluation of Universal Representations of Speech. Covers 102 languages with read speech from Wikipedia passages.
• MLS: Multilingual LibriSpeech. Derived from read audiobooks in 8 languages, providing large-scale multilingual ASR evaluation data.

Metrics

---

• WER (Word Error Rate): Measures the proportion of word-level errors (substitutions, insertions, deletions) in the transcription compared to the reference text. Lower values indicate better accuracy.

Latency Benchmarks

---

We measured RTFx (Real-Time Factor) for each model size on various GPUs. RTFx indicates how many times faster than real-time the model transcribes audio. Higher RTFx is better.

RTFx, batch size 1

RTFx, batched

Benchmarking Methodology

---

The benchmarking was performed on a single GPU using a 10-minute audio file resampled to 16kHz mono. RTFx (Real-Time Factor) is calculated as audio_duration / transcription_time — higher values mean faster-than-real-time transcription.

Algorithm summary:

1. Load the thewhisper-large-v3-turbo model with the specified size (S, M, L, XL, original).
2. Load a 10-minute audio file and resample to 16kHz mono.
3. Run a warm-up pass to initialize GPU caches.
4. Synchronize the GPU, record the start time.
5. Run the transcription pipeline with the specified batch size and chunk length.
6. Synchronize the GPU, record the end time.
7. Calculate RTFx as audio_duration / time_taken.

Serving with Docker Image

---

For serving with Nvidia GPUs, we provide ready-to-go Docker containers with OpenAI-compatible API endpoints. Using our containers you can set up an inference endpoint on any desired cloud/serverless providers as well as on-premise servers. You can also use this container to run inference through TheStage AI platform.

Prebuilt image from ECR

Pull docker image and start inference container. The container serves one model on one GPU; on a multi-GPU host --gpus all is fine, the model is placed on a single device:

docker pull public.ecr.aws/i3f7g5s7/thestage/elastic-models:0.2.4.post0-stt-streaming-v2-20261008-19026a9

docker run --rm -it \
  --name triton-stt \
  --gpus all \
  -p 127.0.0.1:80:80 \
  -v "$HOME/.cache:/opt/project/.cache/" \
  -e MODEL_REPO=TheStageAI/thewhisper-large-v3-turbo \
  -e MODEL_SIZE=<MODEL_SIZE> \
  -e MODEL_BATCH=<MODEL_BATCH> \
  -e PIPELINE_MAX_BATCH_SIZE=<PIPELINE_MAX_BATCH_SIZE> \
  -e CHUNK_LENGTH=<CHUNK_LENGTH> \
  -e PREPROCESSOR_WORKERS=<PREPROCESSOR_WORKERS> \
  -e MODEL_INSTANCES=<MODEL_INSTANCES> \
  -e PREPROCESSOR_QUEUE_DELAY=<PREPROCESSOR_QUEUE_DELAY> \
  -e MODEL_QUEUE_DELAY=<MODEL_QUEUE_DELAY> \
  -e ENSEMBLE_QUEUE_DELAY=<ENSEMBLE_QUEUE_DELAY> \
  -e HUGGINGFACE_ACCESS_TOKEN=<HUGGINGFACE_ACCESS_TOKEN> \
  -e THESTAGE_AUTH_TOKEN=<THESTAGE_ACCESS_TOKEN> \
  public.ecr.aws/i3f7g5s7/thestage/elastic-models:0.2.4.post0-stt-streaming-v2-20261008-19026a9

Invocation

---

CLI

elastic-models-client client stt --sample sample.wav --lang-id en

cURL

curl -X POST http://127.0.0.1:80/v1/audio/transcriptions \
    -H "Authorization: Bearer 123" \
    -H "X-Lang-Id: en" \
    -H "X-Model-Name: thewhisper-large-v3-turbo-<MODEL_SIZE_LOWER>-cl<CHUNK_LENGTH>-bs<MODEL_BATCH>" \
    -F "[email protected]"

Endpoint Parameters

---

Method

POST /v1/audio/transcriptions

Header Parameters

Authorization: string

Bearer token for authentication.

X-Lang-Id: string

Language of the audio (e.g., "en", "es", "fr").

X-Model-Name: string

Specifies the model to use for transcription. Format: thewhisper-large-v3-turbo-<size>-cl<chunk_length>-bs<batch_size>, where <size> is the lowercase letter (s, m, l, xl), <chunk_length> is CHUNK_LENGTH, and <batch_size> is MODEL_BATCH. Example: thewhisper-large-v3-turbo-s-cl15-bs1.

Input Body

file : binary

The audio file to transcribe (multipart/form-data).

Acknowledgments

---

This work builds on Whisper Large V3 Turbo by OpenAI: openai/whisper-large-v3-turbo.

TheWhisper builds on OpenAI Whisper Large V3 Turbo. Optimized and packaged by TheStage AI for ElasticModels / NVIDIA and TheStage Apple SDK.

Links

---

• Original model: openai/whisper-large-v3-turbo
• Platform: app.thestage.ai
• TheStage Apple SDK: github.com/TheStageAI/AppleSDK
• TheWhisper: github.com/TheStageAI/TheWhisper
• Subscribe for updates: TheStageAI X
• Contact email: [email protected]