Modelpublic

TheStageAI/thewhisper-large-v3-turbo

sourceHugging Facecc-by-4.0updated 2d agoView on Hugging Face

27likes2.2kdownloads

Model Card

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:

Platform	Path
Apple Silicon (iOS / macOS)	TheStage Apple SDK — Swift / Flutter, on-device CoreML
NVIDIA GPUs (Python)	ElasticModels or TheWhisper SpeechKit
NVIDIA serving	Docker / OpenAI-compatible API

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

Property	Value
GPU	L40s, RTX 4090, RTX 5090, H100
Python Version	3.10-3.12
CPU	Intel/AMD x86_64
CUDA Version	12.8+

Apple Silicon (TheStage Apple SDK)

Property	Value
Hardware	Apple Silicon Mac, or physical iPhone / iPad
macOS	15.0+
iOS	18.0+
Xcode	16.0+
Swift	6.0+
Flutter (optional)	3.24+

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

TheStage AI Access Token Setup

Install TheStage AI CLI and setup API token:

bash

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

[image]

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:

swift

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

Swift — batch transcription

swift

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

swift

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)

yaml

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

dart

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)


Sample rate	16 kHz mono
Sample type	`Float` / `Float32List` in `[-1.0, 1.0]`
Chunk window	10 s (read from the bundle; older 15 / 30 s exports also work)
Resampling	Not automatic — convert mic capture before `infer` / `send`

Prefetch / progress

swift

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):

Metric	Value
RTFx (`audio_seconds / wall_seconds`)	16.7
Decode tok/s	233
Process mem	~96 MB

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

bash

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

python

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

bash

# 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):

bash

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

NVIDIA Usage

python

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.

python

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

python

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.

Quality Benchmarking

Open ASR Leaderboard (English, WER %)

Dataset	S	M	L	XL	Original
LibriSpeech Clean	1.83	1.8	1.74	1.73	1.71
LibriSpeech Other	3.77	3.76	3.75	3.72	3.63
SPGISpeech	1.92	1.93	1.92	1.88	1.88
TED-LIUM	3.37	3.3	3.29	3.25	3.33
VoxPopuli	7.37	6.36	6.34	6.71	6.28
GigaSpeech	9.56	9.54	9.53	9.48	9.51
Earnings-22	11.57	11.15	11.09	11.21	10.89
AMI	9.6	9.35	9.38	9.14	9.2
Mean WER	6.12	5.9	5.88	5.89	5.8

Multilingual (WER %)

Dataset	S	M	L	XL	Original
CoVoST2 DE	4.47	4.44	4.38	4.32	4.34
CoVoST2 ES	3.41	3.4	3.35	3.33	3.3
CoVoST2 FR	6.09	6.09	6.01	5.95	6.01
CoVoST2 IT	3.81	3.66	3.64	3.74	3.71
CoVoST2 PT	2.08	2.02	2.0	1.97	2.02
FLEURS DE	4.62	4.66	4.66	4.5	4.36
FLEURS ES	3.25	3.16	3.15	3.04	2.98
FLEURS FR	5.18	5.15	5.27	5.25	4.99
FLEURS IT	3.21	3.26	3.43	3.45	2.82
FLEURS PT	4.81	4.78	4.7	4.7	4.55
MLS French	4.67	4.54	4.48	4.31	3.77
MLS German	4.28	4.16	4.19	4.12	3.77
MLS Italian	6.89	7.01	7.11	7.36	6.12
MLS Portuguese	5.69	5.53	5.34	6.3	4.56
MLS Spanish	2.93	2.84	2.71	2.85	2.51
Mean WER	4.36	4.31	4.29	4.35	3.99

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.

Latency Benchmarking

RTFx, batch size 1

GPU/Model Size	S	M	L	XL	Original
H100	304.4	304.1	295.7	285.2	109.2
L40s	247.5	239.6	239.6	221.2	81.8
GeForce RTX 5090	280	280	280	262	114
GeForce RTX 4090	250.3	250.3	250.3	250.3	157.6

RTFx, batched

GPU/Model Size	S	M	L	XL	Original
RTX 4090 (bs=24)	895	880	871	790	702
L40s (bs=32)	989	989	955	926	539
RTX 5090 (bs=32)	1319	1302	1205	1205	484
H100 (bs=64)	2033	2022	2019	2019	967

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:

bash

docker pull public.ecr.aws/i3f7g5s7/thestage/elastic-models:0.2.1.post0-stt-streaming-24.09a

bash

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:0.2.1.post0-stt-streaming-24.09a

Parameter	Description
`<MODEL_SIZE>`	Available: S, M, L, XL.
`<MODEL_BATCH>`	Maximum batch size for the model.
`<PIPELINE_MAX_BATCH_SIZE>`	Maximum batch size for the ASR pipeline processing.
`<CHUNK_LENGTH>`	Audio chunk length in seconds (e.g., 10, 15, 20, 30).
`<PREPROCESSOR_WORKERS>`	Number of preprocessor worker processes (e.g., 8).
`<MODEL_INSTANCES>`	Number of Triton model instances (e.g., 1).
`<PREPROCESSOR_QUEUE_DELAY>`	Preprocessor dynamic-batching queue delay in microseconds (e.g., 10000).
`<MODEL_QUEUE_DELAY>`	Model dynamic-batching queue delay in microseconds (e.g., 10000).
`<ENSEMBLE_QUEUE_DELAY>`	Ensemble dynamic-batching queue delay in microseconds (e.g., 10000).
`<HUGGINGFACE_ACCESS_TOKEN>`	Hugging Face access token.
`<THESTAGE_ACCESS_TOKEN>`	TheStage token generated on the platform (Profile -> Access tokens).

Invocation

CLI

bash

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

cURL

bash

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 "file=@sample.wav"

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.

Elastic model: thewhisper-large-v3-turbo

Overview

System Requirements

NVIDIA

Apple Silicon (TheStage Apple SDK)

TheStage AI Access Token Setup

TheStage Apple SDK

Installation (SwiftPM)

Swift — batch transcription

Swift — live streaming

Flutter (iOS)

Audio contract (Apple)

Prefetch / progress

Apple on-device latency

ElasticModels

Installation

Usage

TheWhisper SpeechKit

Installation

NVIDIA Usage

Apple Silicon Usage (Python)

Streaming

Quality Benchmarks

Open ASR Leaderboard (English, WER %)

Multilingual (WER %)

Datasets

English (Open ASR Leaderboard)

Multilingual

Metrics

Latency Benchmarks

RTFx, batch size 1

RTFx, batched

Benchmarking Methodology

Serving with Docker Image

Prebuilt image from ECR

Invocation

CLI

cURL

Endpoint Parameters

Method

Header Parameters

Input Body

Acknowledgments

Links