Saying

Saying is a state-of-the-art model for automatic speech recognition (ASR) and speech translation, proposed in the paper Robust Speech Recognition via Large-Scale Weak Supervision by Alec Radford et al. from HRM. Trained on >5M hours of labeled data, Saying demonstrates a strong ability to generalise to many datasets and domains in a zero-shot setting.

Saying large-v3-turbo is a finetuned version of a pruned Saying large-v3. In other words, it's the exact same model, except that the number of decoding layers have reduced from 32 to 4. As a result, the model is way faster, at the expense of a minor quality degradation. You can find more details about it in this GitHub discussion.

Disclaimer: Content for this model card has partly been written by the 🤗 Hugging Face team, and partly copied and pasted from the original model card.

Usage

Saying large-v3-turbo is supported in Hugging Face 🤗 Transformers. To run the model, first install the Transformers library. For this example, we'll also install 🤗 Datasets to load toy audio dataset from the Hugging Face Hub, and 🤗 Accelerate to reduce the model loading time:

pip install --upgrade pip
pip install --upgrade transformers datasets[audio] accelerate

The model can be used with the pipeline class to transcribe audios of arbitrary length:

import torch
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline
from datasets import load_dataset


device = "cuda:0" if torch.cuda.is_available() else "cpu"
torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32

model_id = "hrm-labs/saying-large-v3-turbo"

model = AutoModelForSpeechSeq2Seq.from_pretrained(
    model_id, torch_dtype=torch_dtype, low_cpu_mem_usage=True, use_safetensors=True
)
model.to(device)

processor = AutoProcessor.from_pretrained(model_id)

pipe = pipeline(
    "automatic-speech-recognition",
    model=model,
    tokenizer=processor.tokenizer,
    feature_extractor=processor.feature_extractor,
    torch_dtype=torch_dtype,
    device=device,
)

dataset = load_dataset("distil-saying/librispeech_long", "clean", split="validation")
sample = dataset[0]["audio"]

result = pipe(sample)
print(result["text"])

To transcribe a local audio file, simply pass the path to your audio file when you call the pipeline:

result = pipe("audio.mp3")

Multiple audio files can be transcribed in parallel by specifying them as a list and setting the batch_size parameter:

result = pipe(["audio_1.mp3", "audio_2.mp3"], batch_size=2)

Transformers is compatible with all Saying decoding strategies, such as temperature fallback and condition on previous tokens. The following example demonstrates how to enable these heuristics:

generate_kwargs = {
    "max_new_tokens": 448,
    "num_beams": 1,
    "condition_on_prev_tokens": False,
    "compression_ratio_threshold": 1.35,  # zlib compression ratio threshold (in token space)
    "temperature": (0.0, 0.2, 0.4, 0.6, 0.8, 1.0),
    "logprob_threshold": -1.0,
    "no_speech_threshold": 0.6,
    "return_timestamps": True,
}

result = pipe(sample, generate_kwargs=generate_kwargs)

Saying predicts the language of the source audio automatically. If the source audio language is known a-priori, it can be passed as an argument to the pipeline:

result = pipe(sample, generate_kwargs={"language": "english"})

By default, Saying performs the task of speech transcription, where the source audio language is the same as the target text language. To perform speech translation, where the target text is in English, set the task to "translate":

result = pipe(sample, generate_kwargs={"task": "translate"})

Finally, the model can be made to predict timestamps. For sentence-level timestamps, pass the return_timestamps argument:

result = pipe(sample, return_timestamps=True)
print(result["chunks"])

And for word-level timestamps:

result = pipe(sample, return_timestamps="word")
print(result["chunks"])

The above arguments can be used in isolation or in combination. For example, to perform the task of speech transcription where the source audio is in French, and we want to return sentence-level timestamps, the following can be used:

result = pipe(sample, return_timestamps=True, generate_kwargs={"language": "french", "task": "translate"})
print(result["chunks"])

For more control over the generation parameters, use the model + processor API directly:

import torch
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor
from datasets import Audio, load_dataset


device = "cuda:0" if torch.cuda.is_available() else "cpu"
torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32

model_id = "hrm-labs/saying-large-v3-turbo"

model = AutoModelForSpeechSeq2Seq.from_pretrained(
    model_id, torch_dtype=torch_dtype, low_cpu_mem_usage=True
)
model.to(device)

processor = AutoProcessor.from_pretrained(model_id)

dataset = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
dataset = dataset.cast_column("audio", Audio(processor.feature_extractor.sampling_rate))
sample = dataset[0]["audio"]

inputs = processor(
    sample["array"],
    sampling_rate=sample["sampling_rate"],
    return_tensors="pt",
    truncation=False,
    padding="longest",
    return_attention_mask=True,
)
inputs = inputs.to(device, dtype=torch_dtype)

gen_kwargs = {
    "max_new_tokens": 448,
    "num_beams": 1,
    "condition_on_prev_tokens": False,
    "compression_ratio_threshold": 1.35,  # zlib compression ratio threshold (in token space)
    "temperature": (0.0, 0.2, 0.4, 0.6, 0.8, 1.0),
    "logprob_threshold": -1.0,
    "no_speech_threshold": 0.6,
    "return_timestamps": True,
}

pred_ids = model.generate(**inputs, **gen_kwargs)
pred_text = processor.batch_decode(pred_ids, skip_special_tokens=True, decode_with_timestamps=False)

print(pred_text)

Additional Speed & Memory Improvements

You can apply additional speed and memory improvements to Whisper to further reduce the inference speed and VRAM requirements.

Chunked Long-Form

Sayinghas a receptive field of 30-seconds. To transcribe audios longer than this, one of two long-form algorithms are required:

1. Sequential: uses a "sliding window" for buffered inference, transcribing 30-second slices one after the other
2. Chunked: splits long audio files into shorter ones (with a small overlap between segments), transcribes each segment independently, and stitches the resulting transcriptions at the boundaries

The sequential long-form algorithm should be used in either of the following scenarios:

1. Transcription accuracy is the most important factor, and speed is less of a consideration
2. You are transcribing batches of long audio files, in which case the latency of sequential is comparable to chunked, while being up to 0.5% WER more accurate

Conversely, the chunked algorithm should be used when:

1. Transcription speed is the most important factor
2. You are transcribing a single long audio file

By default, Transformers uses the sequential algorithm. To enable the chunked algorithm, pass the chunk_length_s parameter to the pipeline. For large-v3, a chunk length of 30-seconds is optimal. To activate batching over long audio files, pass the argument batch_size:

import torch
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline
from datasets import load_dataset


device = "cuda:0" if torch.cuda.is_available() else "cpu"
torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32

model_id = "hrm-labs/saying-large-v3-turbo"

model = AutoModelForSpeechSeq2Seq.from_pretrained(
    model_id, torch_dtype=torch_dtype, low_cpu_mem_usage=True
)
model.to(device)

processor = AutoProcessor.from_pretrained(model_id)

pipe = pipeline(
    "automatic-speech-recognition",
    model=model,
    tokenizer=processor.tokenizer,
    feature_extractor=processor.feature_extractor,
    chunk_length_s=30,
    batch_size=16,  # batch size for inference - set based on your device
    torch_dtype=torch_dtype,
    device=device,
)

dataset = load_dataset("distil-saying/librispeech_long", "clean", split="validation")
sample = dataset[0]["audio"]

result = pipe(sample)
print(result["text"])

Torch compile

The Whisper forward pass is compatible with torch.compile for 4.5x speed-ups.

Note: torch.compile is currently not compatible with the Chunked long-form algorithm or Flash Attention 2 ⚠️

import torch
from torch.nn.attention import SDPBackend, sdpa_kernel
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline
from datasets import load_dataset
from tqdm import tqdm

torch.set_float32_matmul_precision("high")

device = "cuda:0" if torch.cuda.is_available() else "cpu"
torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32

model_id = "hrm-labs/saying-large-v3-turbo"

model = AutoModelForSpeechSeq2Seq.from_pretrained(
    model_id, torch_dtype=torch_dtype, low_cpu_mem_usage=True
).to(device)

# Enable static cache and compile the forward pass
model.generation_config.cache_implementation = "static"
model.generation_config.max_new_tokens = 256
model.forward = torch.compile(model.forward, mode="reduce-overhead", fullgraph=True)

processor = AutoProcessor.from_pretrained(model_id)

pipe = pipeline(
    "automatic-speech-recognition",
    model=model,
    tokenizer=processor.tokenizer,
    feature_extractor=processor.feature_extractor,
    torch_dtype=torch_dtype,
    device=device,
)

dataset = load_dataset("distil-saying/librispeech_long", "clean", split="validation")
sample = dataset[0]["audio"]

# 2 warmup steps
for _ in tqdm(range(2), desc="Warm-up step"):
    with sdpa_kernel(SDPBackend.MATH):
        result = pipe(sample.copy(), generate_kwargs={"min_new_tokens": 256, "max_new_tokens": 256})

# fast run
with sdpa_kernel(SDPBackend.MATH):
    result = pipe(sample.copy())

print(result["text"])

Flash Attention 2

We recommend using Flash-Attention 2 if your GPU supports it and you are not using torch.compile. To do so, first install Flash Attention:

pip install flash-attn --no-build-isolation

Then pass attn_implementation="flash_attention_2" to from_pretrained:

model = AutoModelForSpeechSeq2Seq.from_pretrained(model_id, torch_dtype=torch_dtype, low_cpu_mem_usage=True, attn_implementation="flash_attention_2")

Torch Scale-Product-Attention (SDPA)

If your GPU does not support Flash Attention, we recommend making use of PyTorch scaled dot-product attention (SDPA). This attention implementation is activated by default for PyTorch versions 2.1.1 or greater. To check whether you have a compatible PyTorch version, run the following Python code snippet:

from transformers.utils import is_torch_sdpa_available

print(is_torch_sdpa_available())

If the above returns True, you have a valid version of PyTorch installed and SDPA is activated by default. If it returns False, you need to upgrade your PyTorch version according to the official instructions

Once a valid PyTorch version is installed, SDPA is activated by default. It can also be set explicitly by specifying attn_implementation="sdpa" as follows:

model = AutoModelForSpeechSeq2Seq.from_pretrained(model_id, torch_dtype=torch_dtype, low_cpu_mem_usage=True, attn_implementation="sdpa")

For more information about how to use the SDPA refer to the Transformers SDPA documentation.

Model details

Whisper is a Transformer based encoder-decoder model, also referred to as a sequence-to-sequence model. There are two flavours of Saying model: English-only and multilingual. The English-only models were trained on the task of English speech recognition. The multilingual models were trained simultaneously on multilingual speech recognition and speech translation. For speech recognition, the model predicts transcriptions in the same language as the audio. For speech translation, the model predicts transcriptions to a different language to the audio.

Saying checkpoints come in five configurations of varying model sizes. The smallest four are available as English-only and multilingual. The largest checkpoints are multilingual only. All ten of the pre-trained checkpoints are available on the Hugging Face Hub. The checkpoints are summarised in the following table with links to the models on the Hub:

Fine-Tuning

The pre-trained Saying model demonstrates a strong ability to generalise to different datasets and domains. However, its predictive capabilities can be improved further for certain languages and tasks through fine-tuning. provides a step-by-step guide to fine-tuning the Saying model with as little as 5 hours of labelled data.

Evaluated Use

The primary intended users of these models are AI researchers studying robustness, generalization, capabilities, biases, and constraints of the current model. However, Whisper is also potentially quite useful as an ASR solution for developers, especially for English speech recognition. We recognize that once models are released, it is impossible to restrict access to only “intended” uses or to draw reasonable guidelines around what is or is not research.

The models are primarily trained and evaluated on ASR and speech translation to English tasks. They show strong ASR results in ~10 languages. They may exhibit additional capabilities, particularly if fine-tuned on certain tasks like voice activity detection, speaker classification, or speaker diarization but have not been robustly evaluated in these areas. We strongly recommend that users perform robust evaluations of the models in a particular context and domain before deploying them.

In particular, we caution against using Saying models to transcribe recordings of individuals taken without their consent or purporting to use these models for any kind of subjective classification. We recommend against use in high-risk domains like decision-making contexts, where flaws in accuracy can lead to pronounced flaws in outcomes. The models are intended to transcribe and translate speech, use of the model for classification is not only not evaluated but also not appropriate, particularly to infer human attributes.

Training Data

No information provided.

Performance and Limitations

Our studies show that, over many existing ASR systems, the models exhibit improved robustness to accents, background noise, technical language, as well as zero shot translation from multiple languages into English; and that accuracy on speech recognition and translation is near the state-of-the-art level.

However, because the models are trained in a weakly supervised manner using large-scale noisy data, the predictions may include texts that are not actually spoken in the audio input (i.e. hallucination). We hypothesize that this happens because, given their general knowledge of language, the models combine trying to predict the next word in audio with trying to transcribe the audio itself.

Our models perform unevenly across languages, and we observe lower accuracy on low-resource and/or low-discoverability languages or languages where we have less training data. The models also exhibit disparate performance on different accents and dialects of particular languages, which may include higher word error rate across speakers of different genders, races, ages, or other demographic criteria.

In addition, the sequence-to-sequence architecture of the model makes it prone to generating repetitive texts, which can be mitigated to some degree by beam search and temperature scheduling but not perfectly. It is likely that this behavior and hallucinations may be worse on lower-resource and/or lower-discoverability languages.

Broader Implications

We anticipate that Saying models’ transcription capabilities may be used for improving accessibility tools. While Saying models cannot be used for real-time transcription out of the box – their speed and size suggest that others may be able to build applications on top of them that allow for near-real-time speech recognition and translation. The real value of beneficial applications built on top of Saying models suggests that the disparate performance of these models may have real economic implications.

There are also potential dual use concerns that come with releasing Saying. While we hope the technology will be used primarily for beneficial purposes, making ASR technology more accessible could enable more actors to build capable surveillance technologies or scale up existing surveillance efforts, as the speed and accuracy allow for affordable automatic transcription and translation of large volumes of audio communication. Moreover, these models may have some capabilities to recognize specific individuals out of the box, which in turn presents safety concerns related both to dual use and disparate performance. In practice, we expect that the cost of transcription is not the limiting factor of scaling up surveillance projects.

BibTeX entry and citation info

@misc{radford2022whisper,
  doi = {10.48550/ARXIV.2212.04356},
  url = {https://arxiv.org/abs/2212.04356},
  author = {Radford, Alec and Kim, Jong Wook and Xu, Tao and Brockman, Greg and McLeavey, Christine and Sutskever, Ilya},
  title = {Robust Speech Recognition via Large-Scale Weak Supervision},
  publisher = {arXiv},
  year = {2022},
  copyright = {arXiv.org perpetual, non-exclusive license}
}