Upload 4 files

run_eval.py

@@ -0,0 +1,241 @@
+import argparse
+
+import io
+import os
+import torch
+import evaluate
+import soundfile
+
+from tqdm import tqdm
+from normalizer import data_utils
+import numpy as np
+
+from nemo.collections.asr.models import ASRModel
+import time
+
+
+wer_metric = evaluate.load("wer")
+
+
+def main(args):
+
+    data_cache_root = args.data_cache_root if args.data_cache_root is not None else os.getcwd()
+    DATA_CACHE_DIR = os.path.join(data_cache_root, "audio_cache")
+    DATASET_NAME = args.dataset
+    SPLIT_NAME = args.split
+
+    CACHE_DIR = os.path.join(DATA_CACHE_DIR, DATASET_NAME, SPLIT_NAME)
+    if not os.path.exists(CACHE_DIR):
+        os.makedirs(CACHE_DIR)
+
+    if args.device >= 0:
+        device = torch.device(f"cuda:{args.device}")
+        compute_dtype=torch.bfloat16
+    else:
+        device = torch.device("cpu")
+        compute_dtype=torch.float32
+
+    if args.model_id.endswith(".nemo"):
+        asr_model = ASRModel.restore_from(args.model_id, map_location=device)
+    else:
+        asr_model = ASRModel.from_pretrained(args.model_id, map_location=device)  # type: ASRModel
+    
+    asr_model.to(compute_dtype)
+    asr_model.eval()
+    print(f"Model size: {sum(p.numel() for p in asr_model.parameters()) / 1e9:.2f}B parameters")
+
+    dataset = data_utils.load_data(args)
+
+    if args.max_eval_samples is not None and args.max_eval_samples > 0:
+        print(f"Subsampling dataset to first {args.max_eval_samples} samples !")
+        dataset = dataset.take(args.max_eval_samples)
+
+    # Prepare data FIRST - this casts audio to proper format with "array" and "sampling_rate" keys
+    dataset = data_utils.prepare_data(dataset)
+
+    def download_audio_files(batch):
+
+        # download audio files and write the paths, transcriptions and durations to a manifest file
+        audio_paths = []
+        original_audio_paths = []
+        durations = []
+        file_names = batch.get("file_name", [None] * len(batch["audio"]))
+
+        # Use 'id' column if available, otherwise generate sequential IDs
+        if "id" in batch:
+            ids = batch["id"]
+        else:
+            # Generate IDs based on index
+            start_idx = len([f for f in os.listdir(CACHE_DIR) if f.endswith('.wav')]) if os.path.exists(CACHE_DIR) else 0
+            ids = [f"sample_{start_idx + i}" for i in range(len(batch["audio"]))]
+
+        for id, file_name, audio_sample in zip(ids, file_names, batch["audio"]):
+
+            # first step added here to make ID and wav filenames unique
+            # several datasets like earnings22 have a hierarchical structure
+            # for eg. earnings22/test/4432298/281.wav, earnings22/test/4450488/281.wav
+            # lhotse uses the filename (281.wav) here as unique ID to create and name cuts
+            # ref: https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/collation.py#L186
+            original_id = id  # preserve before sanitization for use as audio_filepath
+            id = id.replace('/', '_').removesuffix('.wav')
+
+            audio_path = os.path.join(CACHE_DIR, f"{id}.wav")
+            audio_array = np.float32(audio_sample["array"])
+            sample_rate = audio_sample["sampling_rate"]
+
+            if not os.path.exists(audio_path):
+                os.makedirs(os.path.dirname(audio_path), exist_ok=True)
+                soundfile.write(audio_path, audio_array, sample_rate)
+
+            audio_paths.append(audio_path)
+            # Prefer the original file_name from the dataset; fall back to the
+            # sample id (before path-sanitization) so audio_filepath in the
+            # JSONL is always a meaningful identifier rather than "sample_N".
+            if file_name is not None:
+                original_audio_paths.append(os.path.basename(str(file_name)))
+            else:
+                original_audio_paths.append(original_id)
+            durations.append(len(audio_array) / sample_rate)
+
+        
+        batch["references"] = batch["norm_text"]
+        batch["audio_filepaths"] = audio_paths
+        batch["original_audio_filepaths"] = original_audio_paths
+        batch["durations"] = durations
+
+        return batch
+
+    if asr_model.cfg.decoding.strategy != "beam":
+        asr_model.cfg.decoding.strategy = "greedy_batch"
+        asr_model.change_decoding_strategy(asr_model.cfg.decoding)
+    
+    # prepraing the offline dataset
+    dataset = dataset.map(download_audio_files, batch_size=args.batch_size, batched=True, remove_columns=["audio"])
+
+    # Write manifest from daraset batch using json and keys audio_filepath, duration, text
+
+    all_data = {
+        "audio_filepaths": [],
+        "original_audio_filepaths": [],
+        "durations": [],
+        "references": [],
+    }
+
+    data_itr = iter(dataset)
+    for data in tqdm(data_itr, desc="Downloading Samples"):
+        for key in all_data:
+            all_data[key].append(data[key])
+    
+    # Sort audio_filepaths and references based on durations values
+    sorted_indices = sorted(range(len(all_data["durations"])), key=lambda k: all_data["durations"][k], reverse=True)
+    all_data["audio_filepaths"] = [all_data["audio_filepaths"][i] for i in sorted_indices]
+    all_data["original_audio_filepaths"] = [all_data["original_audio_filepaths"][i] for i in sorted_indices]
+    all_data["references"] = [all_data["references"][i] for i in sorted_indices]
+    all_data["durations"] = [all_data["durations"][i] for i in sorted_indices]
+    
+    
+    total_time = 0
+    for _ in range(2): # warmup once and calculate rtf
+        if _ == 0:
+            audio_files = all_data["audio_filepaths"][:args.batch_size * 4] # warmup with 4 batches
+        else:
+            audio_files = all_data["audio_filepaths"]
+        start_time = time.time()
+        with torch.inference_mode(), torch.no_grad(): 
+
+            if 'canary' in args.model_id and 'v2' not in args.model_id:
+                pnc = 'nopnc'
+            else:
+                pnc = 'pnc'
+                
+            if 'canary' in args.model_id:
+                transcriptions = asr_model.transcribe(audio_files, batch_size=args.batch_size, verbose=False, pnc=pnc, num_workers=1)
+            else:
+                transcriptions = asr_model.transcribe(audio_files, batch_size=args.batch_size, verbose=False, num_workers=1)
+        end_time = time.time()
+        if _ == 1:
+            total_time += end_time - start_time
+    total_time = total_time
+
+    # normalize transcriptions with English normalizer
+    if isinstance(transcriptions, tuple) and len(transcriptions) == 2:
+        transcriptions = transcriptions[0]
+    predictions = [data_utils.normalizer(pred.text) for pred in transcriptions]
+
+    avg_time = total_time / len(all_data["audio_filepaths"])
+
+    # Write manifest results (WER and RTFX)
+    manifest_path = data_utils.write_manifest(
+        all_data["references"],
+        predictions,
+        args.model_id,
+        args.dataset_path,
+        args.dataset,
+        args.split,
+        audio_length=all_data["durations"],
+        transcription_time=[avg_time] * len(all_data["audio_filepaths"]),
+        audio_filepaths=all_data["original_audio_filepaths"],
+    )
+
+    print("Results saved at path:", os.path.abspath(manifest_path))
+
+    wer = wer_metric.compute(references=all_data['references'], predictions=predictions)
+    wer = round(100 * wer, 2)
+
+    # transcription_time = sum(all_results["transcription_time"])
+    audio_length = sum(all_data["durations"])
+    rtfx = audio_length / total_time
+    rtfx = round(rtfx, 2)
+
+    print("RTFX:", rtfx)
+    print("WER:", wer, "%")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--model_id", type=str, required=True, help="Model identifier. Should be loadable with NVIDIA NeMo.",
+    )
+    parser.add_argument(
+        '--dataset_path', type=str, default='hf-audio/open-asr-leaderboard', help='Dataset path. By default, it is `hf-audio/open-asr-leaderboard`'
+    )
+    parser.add_argument(
+        '--data_cache_root', type=str, default=None, help='Root directory for audio cache. By default, it is the current working directory.'
+    )
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        required=True,
+        help="Dataset name. *E.g.* `'librispeech_asr` for the LibriSpeech ASR dataset, or `'common_voice'` for Common Voice. The full list of dataset names "
+        "can be found at `https://huggingface.co/datasets/hf-audio/open-asr-leaderboard`",
+    )
+    parser.add_argument(
+        "--split",
+        type=str,
+        default="test",
+        help="Split of the dataset. *E.g.* `'validation`' for the dev split, or `'test'` for the test split.",
+    )
+    parser.add_argument(
+        "--device",
+        type=int,
+        default=-1,
+        help="The device to run the pipeline on. -1 for CPU (default), 0 for the first GPU and so on.",
+    )
+    parser.add_argument(
+        "--batch_size", type=int, default=32, help="Number of samples to go through each streamed batch.",
+    )
+    parser.add_argument(
+        "--max_eval_samples",
+        type=int,
+        default=None,
+        help="Number of samples to be evaluated. Put a lower number e.g. 64 for testing this script.",
+    )
+    parser.add_argument(
+        "--streaming",
+        action="store_true",
+        help="Stream the dataset lazily over the network instead of downloading it in full before the evaluation. Off by default for reproducible benchmark timings.",
+    )
+    args = parser.parse_args()
+
+    main(args)

run_eval_long.py

@@ -0,0 +1,224 @@
+import argparse
+
+import io
+import os
+import torch
+import evaluate
+import soundfile
+
+from tqdm import tqdm
+from normalizer import data_utils
+import numpy as np
+
+from nemo.collections.asr.models import ASRModel
+import time
+
+
+wer_metric = evaluate.load("wer")
+
+
+def main(args):
+
+    DATA_CACHE_DIR = os.path.join(os.getcwd(), "audio_cache")
+    DATASET_NAME = args.dataset
+    SPLIT_NAME = args.split
+
+    CACHE_DIR = os.path.join(DATA_CACHE_DIR, DATASET_NAME, SPLIT_NAME)
+    if not os.path.exists(CACHE_DIR):
+        os.makedirs(CACHE_DIR)
+
+    if args.device >= 0:
+        device = torch.device(f"cuda:{args.device}")
+        compute_dtype=torch.bfloat16
+    else:
+        device = torch.device("cpu")
+        compute_dtype=torch.float32
+        
+
+    if args.model_id.endswith(".nemo"):
+        asr_model = ASRModel.restore_from(args.model_id, map_location=device)
+    else:
+        asr_model = ASRModel.from_pretrained(args.model_id, map_location=device)  # type: ASRModel
+       
+    if args.longform:
+        asr_model.change_attention_model("rel_pos_local_attn", [128, 128])  # local attn
+    asr_model.to(compute_dtype)
+    asr_model.eval()
+    print(f"Model size: {sum(p.numel() for p in asr_model.parameters()) / 1e9:.2f}B parameters")
+
+    dataset = data_utils.load_data(args)
+
+    def download_audio_files(batch, indices):
+
+        # download audio files and write the paths, transcriptions and durations to a manifest file
+        audio_paths = []
+        durations = []
+
+        # Use global indices for unique filenames across all batches
+        for global_idx, sample in zip(indices, batch["audio"]):
+            # Use a unique filename based on global index
+            audio_path = os.path.join(CACHE_DIR, f"sample_{global_idx}.wav")
+
+            if "array" in sample:
+                audio_array = np.float32(sample["array"])
+                sample_rate = 16000
+
+            elif "bytes" in sample: # added to be compatible with latest datasets library (3.x.x) that produces byte stream
+                with io.BytesIO(sample["bytes"]) as audio_file:
+                    audio_array, sample_rate = soundfile.read(audio_file, dtype="float32")
+
+            else:
+                raise ValueError("Sample must have either 'array' or 'bytes' key")
+
+            if not os.path.exists(audio_path):
+                os.makedirs(os.path.dirname(audio_path), exist_ok=True)
+                soundfile.write(audio_path, audio_array, sample_rate)
+
+            audio_paths.append(audio_path)
+            durations.append(len(audio_array) / sample_rate)
+
+        
+        batch["references"] = batch["norm_text"]
+        batch["audio_filepaths"] = audio_paths
+        batch["durations"] = durations
+
+        return batch
+
+
+    if args.max_eval_samples is not None and args.max_eval_samples > 0:
+        print(f"Subsampling dataset to first {args.max_eval_samples} samples !")
+        dataset = dataset.take(args.max_eval_samples)
+
+    dataset = data_utils.prepare_data(dataset)
+    if asr_model.cfg.decoding.strategy != "beam":
+        asr_model.cfg.decoding.strategy = "greedy_batch"
+        asr_model.change_decoding_strategy(asr_model.cfg.decoding)
+    
+    # prepraing the offline dataset
+    dataset = dataset.map(download_audio_files, batch_size=args.batch_size, batched=True, with_indices=True, remove_columns=["audio"])
+
+    # Write manifest from daraset batch using json and keys audio_filepath, duration, text
+
+    all_data = {
+        "audio_filepaths": [],
+        "durations": [],
+        "references": [],
+    }
+
+    data_itr = iter(dataset)
+    for data in tqdm(data_itr, desc="Downloading Samples"):
+        for key in all_data:
+            all_data[key].append(data[key])
+    
+    # Sort audio_filepaths and references based on durations values
+    sorted_indices = sorted(range(len(all_data["durations"])), key=lambda k: all_data["durations"][k], reverse=True)
+    all_data["audio_filepaths"] = [all_data["audio_filepaths"][i] for i in sorted_indices]
+    all_data["references"] = [all_data["references"][i] for i in sorted_indices]
+    all_data["durations"] = [all_data["durations"][i] for i in sorted_indices]
+
+    total_time = 0
+    for _ in range(2): # warmup once and calculate rtf
+        if _ == 0:
+            audio_files = all_data["audio_filepaths"][:args.batch_size * 4] # warmup with 4 batches
+        else:
+            audio_files = all_data["audio_filepaths"]
+        start_time = time.time()
+        with torch.inference_mode(), torch.no_grad(): 
+
+            if 'canary' in args.model_id and 'v2' not in args.model_id:
+                pnc = 'nopnc'
+            else:
+                pnc = 'pnc'
+                
+            if 'canary' in args.model_id:
+                transcriptions = asr_model.transcribe(audio_files, batch_size=args.batch_size, verbose=False, pnc=pnc, num_workers=1)
+            else:
+                transcriptions = asr_model.transcribe(audio_files, batch_size=args.batch_size, verbose=False, num_workers=1)
+        end_time = time.time()
+        if _ == 1:
+            total_time += end_time - start_time
+    total_time = total_time
+
+    # normalize transcriptions with English normalizer
+    if isinstance(transcriptions, tuple) and len(transcriptions) == 2:
+        transcriptions = transcriptions[0]
+    predictions = [data_utils.normalizer(pred.text) for pred in transcriptions]
+
+    avg_time = total_time / len(all_data["audio_filepaths"])
+
+    # Write manifest results (WER and RTFX)
+    manifest_path = data_utils.write_manifest(
+        all_data["references"],
+        predictions,
+        args.model_id,
+        args.dataset_path,
+        args.dataset,
+        args.split,
+        audio_length=all_data["durations"],
+        transcription_time=[avg_time] * len(all_data["audio_filepaths"]),
+    )
+
+    print("Results saved at path:", os.path.abspath(manifest_path))
+
+    wer = wer_metric.compute(references=all_data['references'], predictions=predictions)
+    wer = round(100 * wer, 2)
+
+    # transcription_time = sum(all_results["transcription_time"])
+    audio_length = sum(all_data["durations"])
+    rtfx = audio_length / total_time
+    rtfx = round(rtfx, 2)
+
+    print("RTFX:", rtfx)
+    print("WER:", wer, "%")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--model_id", type=str, required=True, help="Model identifier. Should be loadable with NVIDIA NeMo.",
+    )
+    parser.add_argument(
+        '--dataset_path', type=str, default='hf-audio/open-asr-leaderboard', help='Dataset path. By default, it is `hf-audio/open-asr-leaderboard`'
+    )
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        required=True,
+        help="Dataset name. *E.g.* `'librispeech_asr` for the LibriSpeech ASR dataset, or `'common_voice'` for Common Voice. The full list of dataset names "
+        "can be found at `https://huggingface.co/datasets/hf-audio/open-asr-leaderboard`",
+    )
+    parser.add_argument(
+        "--split",
+        type=str,
+        default="test",
+        help="Split of the dataset. *E.g.* `'validation`' for the dev split, or `'test'` for the test split.",
+    )
+    parser.add_argument(
+        "--device",
+        type=int,
+        default=-1,
+        help="The device to run the pipeline on. -1 for CPU (default), 0 for the first GPU and so on.",
+    )
+    parser.add_argument(
+        "--batch_size", type=int, default=32, help="Number of samples to go through each streamed batch.",
+    )
+    parser.add_argument(
+        "--max_eval_samples",
+        type=int,
+        default=None,
+        help="Number of samples to be evaluated. Put a lower number e.g. 64 for testing this script.",
+    )
+    parser.add_argument(
+        "--streaming",
+        action="store_true",
+        help="Stream the dataset lazily over the network instead of downloading it in full before the evaluation. Off by default for reproducible benchmark timings.",
+    )
+    parser.add_argument(
+        "--longform",
+        action="store_true",
+        help="Whether to use longform mode.",
+    )
+    args = parser.parse_args()
+
+    main(args)

run_eval_ml.py

@@ -0,0 +1,280 @@
+# This script is used to evaluate NeMo ASR models on the Multi-Lingual datasets
+
+import argparse
+import io
+import os
+os.environ["DATASETS_USE_TORCHCODEC"] = "0"
+import torch
+import evaluate
+import soundfile
+import numpy as np
+from tqdm import tqdm
+from datasets import load_dataset
+from normalizer import data_utils
+from normalizer.eval_utils import normalize_compound_pairs
+from nemo.collections.asr.models import ASRModel
+from omegaconf import OmegaConf
+import time
+
+
+wer_metric = evaluate.load("wer")
+
+
+def main(args):
+    DATA_CACHE_DIR = os.path.join(os.getcwd(), "audio_cache")
+    CONFIG_NAME = args.config_name
+    SPLIT_NAME = args.split
+    
+    # Extract language from config_name if not provided
+    if args.language:
+        LANGUAGE = args.language
+    else:
+        # Extract language from config_name (e.g., "fleurs_en" -> "en")
+        try:
+            LANGUAGE = CONFIG_NAME.split('_', 1)[1]
+        except IndexError:
+            LANGUAGE = "en"  # Default fallback
+    
+    print(f"Detected language: {LANGUAGE}")
+
+    CACHE_DIR = os.path.join(DATA_CACHE_DIR, CONFIG_NAME, SPLIT_NAME)
+    if not os.path.exists(CACHE_DIR):
+        os.makedirs(CACHE_DIR)
+
+    if args.device >= 0:
+        device = torch.device(f"cuda:{args.device}")
+        compute_dtype = torch.bfloat16
+    else:
+        device = torch.device("cpu")
+        compute_dtype = torch.float32
+
+    # Load ASR model
+    if args.model_id.endswith(".nemo"):
+        asr_model = ASRModel.restore_from(args.model_id, map_location=device)
+    else:
+        asr_model = ASRModel.from_pretrained(args.model_id, map_location=device)
+    
+    asr_model.to(compute_dtype)
+    asr_model.eval()
+    print(f"Model size: {sum(p.numel() for p in asr_model.parameters()) / 1e9:.2f}B parameters")
+
+    # Load dataset using the HuggingFace dataset repository
+    print(f"Loading dataset: {args.dataset} with config: {CONFIG_NAME}")
+
+    dataset = load_dataset(args.dataset, CONFIG_NAME, split=SPLIT_NAME, streaming=args.streaming)
+    
+    if args.max_eval_samples is not None and args.max_eval_samples > 0:
+        print(f"Subsampling dataset to first {args.max_eval_samples} samples!")
+        dataset = dataset.select(range(min(args.max_eval_samples, len(dataset))))
+
+    # Configure decoding strategy
+    if asr_model.cfg.decoding.strategy != "beam":
+        asr_model.cfg.decoding.strategy = "greedy_batch"
+        if hasattr(asr_model.cfg.decoding, "greedy"):
+            OmegaConf.update(asr_model.cfg.decoding, "greedy.use_cuda_graph_decoder", False, force_add=True)
+        asr_model.change_decoding_strategy(asr_model.cfg.decoding)
+
+    def download_audio_files(batch):
+        """Process audio files and prepare them for evaluation."""
+        audio_paths = []
+        durations = []
+
+        for i, (file_name, sample, duration, text) in enumerate(zip(
+            batch["file_name"], batch["audio"], batch["duration"], batch["text"]
+        )):
+            # Create unique filename using index to avoid conflicts
+            unique_id = f"{CONFIG_NAME}_{i}_{os.path.basename(file_name).replace('.wav', '')}"
+            audio_path = os.path.join(CACHE_DIR, f"{unique_id}.wav")
+
+            if "array" in sample:
+                audio_array = np.float32(sample["array"])
+                sample_rate = sample.get("sampling_rate", 16000)
+            elif "bytes" in sample:
+                with io.BytesIO(sample["bytes"]) as audio_file:
+                    audio_array, sample_rate = soundfile.read(audio_file, dtype="float32")
+            else:
+                raise ValueError("Sample must have either 'array' or 'bytes' key")
+
+            if not os.path.exists(audio_path):
+                os.makedirs(os.path.dirname(audio_path), exist_ok=True)
+                soundfile.write(audio_path, audio_array, sample_rate)
+
+            audio_paths.append(audio_path)
+            # Use duration from dataset if available, otherwise calculate
+            if duration is not None:
+                durations.append(duration)
+            else:
+                durations.append(len(audio_array) / sample_rate)
+
+        batch["references"] = [text for text in batch["text"]]
+        batch["audio_filepaths"] = audio_paths
+        batch["durations"] = durations
+
+        return batch
+
+    # Process the dataset
+    print("Processing audio files...")
+    dataset = dataset.map(
+        download_audio_files, 
+        batch_size=args.batch_size, 
+        batched=True, 
+        remove_columns=["audio"]
+    )
+
+    # Collect all data
+    all_data = {
+        "audio_filepaths": [],
+        "durations": [],
+        "references": [],
+    }
+
+    print("Collecting data...")
+    for data in tqdm(dataset, desc="Collecting samples"):
+        all_data["audio_filepaths"].append(data["audio_filepaths"])
+        all_data["durations"].append(data["durations"])
+        all_data["references"].append(data["references"])
+
+    # Sort by duration for efficient batch processing
+    print("Sorting by duration...")
+    sorted_indices = sorted(range(len(all_data["durations"])), key=lambda k: all_data["durations"][k], reverse=True)
+    all_data["audio_filepaths"] = [all_data["audio_filepaths"][i] for i in sorted_indices]
+    all_data["references"] = [all_data["references"][i] for i in sorted_indices]
+    all_data["durations"] = [all_data["durations"][i] for i in sorted_indices]
+
+    # Run evaluation with warmup
+    total_time = 0
+    for warmup_round in range(2):  # warmup once and calculate rtf
+        if warmup_round == 0:
+            audio_files = all_data["audio_filepaths"][:args.batch_size * 4]  # warmup with 4 batches
+            print("Running warmup...")
+        else:
+            audio_files = all_data["audio_filepaths"]
+            print("Running full evaluation...")
+            
+        start_time = time.time()
+        with torch.inference_mode(), torch.no_grad():
+            # for canary-1b and canary-1b-flash, we need to set pnc='no' for English and for other languages, we need to set pnc='pnc' but for canary-1b-v2 pnc='yes' for all languages
+            if 'canary' in args.model_id and 'v2' not in args.model_id:
+                pnc = 'nopnc' if LANGUAGE == "en" else 'pnc'
+            else:
+                pnc = 'pnc'
+
+            if 'canary' in args.model_id:
+                transcriptions = asr_model.transcribe(audio_files, batch_size=args.batch_size, verbose=False, pnc=pnc, num_workers=1, source_lang=LANGUAGE, target_lang=LANGUAGE)
+            else:
+                transcriptions = asr_model.transcribe(audio_files, batch_size=args.batch_size, verbose=False, num_workers=1)
+        end_time = time.time()
+        
+        if warmup_round == 1:
+            total_time = end_time - start_time
+
+    # Process transcriptions
+    if isinstance(transcriptions, tuple) and len(transcriptions) == 2:
+        transcriptions = transcriptions[0]
+    
+    references = all_data["references"] 
+    if LANGUAGE == "en": # English is handled by the English normalizer
+        references = [data_utils.normalizer(ref) for ref in references]
+        predictions = [data_utils.normalizer(pred.text) for pred in transcriptions]
+    else:
+        references = [data_utils.ml_normalizer(ref, lang=LANGUAGE) for ref in references]
+        predictions = [data_utils.ml_normalizer(pred.text, lang=LANGUAGE) for pred in transcriptions]
+
+    # Filter empty references (consistent with English pipeline)
+    filtered = [
+        (ref, pred, dur)
+        for ref, pred, dur in zip(references, predictions, all_data["durations"])
+        if data_utils.is_target_text_in_range(ref)
+    ]
+    if filtered:
+        references, predictions, all_data["durations"] = zip(*filtered)
+        references, predictions = list(references), list(predictions)
+        all_data["durations"] = list(all_data["durations"])
+
+    avg_time = total_time / len(all_data["audio_filepaths"])
+
+    # Write results using eval_utils.write_manifest
+    manifest_path = data_utils.write_manifest(
+        references,
+        predictions,
+        args.model_id,
+        args.dataset,  # dataset_path for filename
+        CONFIG_NAME,  # dataset_name
+        SPLIT_NAME,
+        audio_length=all_data["durations"],
+        transcription_time=[avg_time] * len(all_data["audio_filepaths"]),
+    )
+
+    print("Results saved at path:", os.path.abspath(manifest_path))
+
+    # Calculate metrics
+    wer_refs, wer_preds = normalize_compound_pairs(references, predictions)
+    wer = wer_metric.compute(references=wer_refs, predictions=wer_preds)
+    wer = round(100 * wer, 2)
+
+    audio_length = sum(all_data["durations"])
+    rtfx = audio_length / total_time
+    rtfx = round(rtfx, 2)
+
+    print(f"Dataset: {args.dataset}")
+    print(f"Language: {LANGUAGE}")
+    print(f"Config: {CONFIG_NAME}")
+    print(f"Model: {args.model_id}")
+    print(f"RTFX: {rtfx}")
+    print(f"WER: {wer}%")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--model_id", type=str, required=True, help="Model identifier. Should be loadable with NVIDIA NeMo.",
+    )
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        default="nithinraok/asr-leaderboard-datasets",
+        help="Dataset name. Default is 'nithinraok/asr-leaderboard-datasets'"
+    )
+    parser.add_argument(
+        "--config_name",
+        type=str,
+        required=True,
+        help="Config name in format <dataset>_<lang> (e.g., fleurs_en, mcv_de, mls_es)"
+    )
+    parser.add_argument(
+        "--language",
+        type=str,
+        default=None,
+        help="Language code (e.g., en, de, es). If not provided, will be extracted from config_name."
+    )
+    parser.add_argument(
+        "--split",
+        type=str,
+        default="test",
+        help="Split of the dataset. Default is 'test'.",
+    )
+    parser.add_argument(
+        "--device",
+        type=int,
+        default=-1,
+        help="The device to run the pipeline on. -1 for CPU (default), 0 for the first GPU and so on.",
+    )
+    parser.add_argument(
+        "--batch_size", type=int, default=32, help="Number of samples to go through each streamed batch.",
+    )
+    parser.add_argument(
+        "--max_eval_samples",
+        type=int,
+        default=None,
+        help="Number of samples to be evaluated. Put a lower number e.g. 64 for testing this script.",
+    )
+
+    parser.add_argument(
+        "--streaming",
+        action="store_true",
+        help="Stream the dataset lazily over the network instead of downloading it in full before the evaluation. Off by default for reproducible benchmark timings.",
+    )
+    args = parser.parse_args()
+
+    main(args) 

run_eval_salm.py

@@ -0,0 +1,276 @@
+import argparse
+
+import io
+import os
+import torch
+import evaluate
+import soundfile
+import lhotse
+
+from tqdm import tqdm
+from normalizer import data_utils
+import numpy as np
+
+from nemo.collections.asr.models import ASRModel
+import time
+
+
+from nemo.collections.speechlm2.models.salm import SALM
+from omegaconf import OmegaConf
+from pathlib import Path
+from transformers import GenerationConfig
+
+
+
+wer_metric = evaluate.load("wer")
+
+
+class ToAudio(torch.utils.data.Dataset):
+    def __getitem__(self, cuts):
+        cuts = lhotse.CutSet([c.to_mono(mono_downmix=True) if isinstance(c, lhotse.MultiCut) else c for c in cuts])
+        audios, audio_lens = cuts.load_audio(collate=True)
+        return {"cuts": cuts, "audios": audios, "audio_lens": audio_lens}
+
+
+def setup_dloader(audio_files, batch_size, num_workers):
+    cuts = lhotse.CutSet([lhotse.Recording.from_file(p).to_cut() for p in audio_files])
+    cuts = cuts.resample(16000)
+    return torch.utils.data.DataLoader(
+        dataset=ToAudio(),
+        sampler=lhotse.dataset.DynamicCutSampler(cuts, max_cuts=batch_size),
+        num_workers=num_workers,
+        batch_size=None,
+    )
+
+
+def transcribe(model, dloader) -> list[str]:
+    hyps = []
+    eos_tokens = torch.tensor([model.text_eos_id])
+    for batch_idx, batch in enumerate(dloader):
+        answer_ids = model.generate(
+            prompts=[
+                [
+                    {"role": "user", "slots": {"message": f"Transcribe the following: {model.audio_locator_tag}"}}
+                ]
+            ] * len(batch["cuts"]),
+            audios=batch["audios"].to(model.device, non_blocking=True),
+            audio_lens=batch["audio_lens"].to(model.device, non_blocking=True),
+            generation_config=GenerationConfig(
+                max_new_tokens=128,
+                bos_token_id=model.text_bos_id,
+                eos_token_id=eos_tokens,
+                pad_token_id=model.text_pad_id,
+            ),
+        )
+        answer_ids = [parse_hyp(ans, eos_tokens) for ans in answer_ids.cpu()]
+        hyps.extend(model.tokenizer.ids_to_text(ans).strip() for ans in answer_ids)
+    return hyps
+
+
+def parse_hyp(answer: torch.Tensor, eos_tokens):
+    end = (answer == torch.isin(answer, eos_tokens)).nonzero(as_tuple=True)[0]
+    if end.numel() == 0:
+        return answer
+    end = end[0]
+    return answer[:end]
+
+
+def main(args):
+
+    data_cache_root = args.data_cache_root if args.data_cache_root is not None else os.getcwd()
+    DATA_CACHE_DIR = os.path.join(data_cache_root, "audio_cache")
+    DATASET_NAME = args.dataset
+    SPLIT_NAME = args.split
+
+    CACHE_DIR = os.path.join(DATA_CACHE_DIR, DATASET_NAME, SPLIT_NAME)
+    if not os.path.exists(CACHE_DIR):
+        os.makedirs(CACHE_DIR)
+
+    torch.set_float32_matmul_precision("medium")
+
+    device = torch.device(f"cuda:{args.device}")
+    model = SALM.from_pretrained(args.model_id).eval().to(torch.bfloat16).to(device)
+    print(f"Model size: {sum(p.numel() for p in model.parameters()) / 1e9:.2f}B parameters")
+
+    dataset = data_utils.load_data(args)
+
+    def download_audio_files(batch):
+
+        # download audio files and write the paths, transcriptions and durations to a manifest file
+        audio_paths = []
+        original_audio_paths = []
+        durations = []
+        file_names = batch.get("file_name", [None] * len(batch["audio"]))
+
+        # Use 'id' column if available, otherwise generate sequential IDs
+        if "id" in batch:
+            ids = batch["id"]
+        else:
+            # Generate IDs based on index
+            start_idx = len([f for f in os.listdir(CACHE_DIR) if f.endswith('.wav')]) if os.path.exists(CACHE_DIR) else 0
+            ids = [f"sample_{start_idx + i}" for i in range(len(batch["audio"]))]
+
+        for id, file_name, sample in zip(ids, file_names, batch["audio"]):
+
+            # first step added here to make ID and wav filenames unique
+            # several datasets like earnings22 have a hierarchical structure
+            # for eg. earnings22/test/4432298/281.wav, earnings22/test/4450488/281.wav
+            # lhotse uses the filename (281.wav) here as unique ID to create and name cuts
+            # ref: https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/collation.py#L186
+            original_id = id  # preserve before sanitization for use as audio_filepath
+            id = id.replace('/', '_').removesuffix('.wav')
+
+            audio_path = os.path.join(CACHE_DIR, f"{id}.wav")
+            audio_array = np.float32(sample["array"])
+            sample_rate = sample["sampling_rate"]
+
+            if not os.path.exists(audio_path):
+                os.makedirs(os.path.dirname(audio_path), exist_ok=True)
+                soundfile.write(audio_path, audio_array, sample_rate)
+
+            audio_paths.append(audio_path)
+            if file_name is not None:
+                original_audio_paths.append(os.path.basename(str(file_name)))
+            else:
+                original_audio_paths.append(original_id)
+            durations.append(len(audio_array) / sample_rate)
+
+
+        batch["references"] = batch["norm_text"]
+        batch["audio_filepaths"] = audio_paths
+        batch["original_audio_filepaths"] = original_audio_paths
+        batch["durations"] = durations
+
+        return batch
+
+
+    if args.max_eval_samples is not None and args.max_eval_samples > 0:
+        print(f"Subsampling dataset to first {args.max_eval_samples} samples !")
+        dataset = dataset.take(args.max_eval_samples)
+
+    dataset = data_utils.prepare_data(dataset)
+
+    # prepraing the offline dataset
+    dataset = dataset.map(download_audio_files, batch_size=args.batch_size, batched=True, remove_columns=["audio"])
+
+    # Write manifest from daraset batch using json and keys audio_filepath, duration, text
+
+    all_data = {
+        "audio_filepaths": [],
+        "original_audio_filepaths": [],
+        "durations": [],
+        "references": [],
+    }
+
+    data_itr = iter(dataset)
+    for data in tqdm(data_itr, desc="Downloading Samples"):
+        for key in all_data:
+            all_data[key].append(data[key])
+
+    # Sort audio_filepaths and references based on durations values
+    sorted_indices = sorted(range(len(all_data["durations"])), key=lambda k: all_data["durations"][k], reverse=True)
+    all_data["audio_filepaths"] = [all_data["audio_filepaths"][i] for i in sorted_indices]
+    all_data["original_audio_filepaths"] = [all_data["original_audio_filepaths"][i] for i in sorted_indices]
+    all_data["references"] = [all_data["references"][i] for i in sorted_indices]
+    all_data["durations"] = [all_data["durations"][i] for i in sorted_indices]
+
+
+    total_time = 0
+    for _ in range(2): # warmup once and calculate rtf
+        if _ == 0:
+            audio_files = all_data["audio_filepaths"][:args.batch_size * 4] # warmup with 4 batches
+        else:
+            audio_files = all_data["audio_filepaths"]
+        dloader = setup_dloader(audio_files=audio_files, batch_size=args.batch_size, num_workers=1)
+        with torch.inference_mode():
+            start_time = time.time()
+            transcriptions = transcribe(model, dloader)
+            end_time = time.time()
+        if _ == 1:
+            total_time += end_time - start_time
+    total_time = total_time
+
+    # normalize transcriptions with English normalizer
+    if isinstance(transcriptions, tuple) and len(transcriptions) == 2:
+        transcriptions = transcriptions[0]
+    predictions = [data_utils.normalizer(pred) for pred in transcriptions]
+
+    avg_time = total_time / len(all_data["audio_filepaths"])
+
+    # Write manifest results (WER and RTFX)
+    manifest_path = data_utils.write_manifest(
+        all_data["references"],
+        predictions,
+        args.model_id,
+        args.dataset_path,
+        args.dataset,
+        args.split,
+        audio_length=all_data["durations"],
+        transcription_time=[avg_time] * len(all_data["audio_filepaths"]),
+        audio_filepaths=all_data["original_audio_filepaths"],
+    )
+
+    print("Results saved at path:", os.path.abspath(manifest_path))
+
+    wer = wer_metric.compute(references=all_data['references'], predictions=predictions)
+    wer = round(100 * wer, 2)
+
+    audio_length = sum(all_data["durations"])
+    rtfx = audio_length / total_time
+    rtfx = round(rtfx, 2)
+
+    print("RTFX:", rtfx)
+    print("WER:", wer, "%")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--model_id", type=str, required=True, help="Model identifier. Should be loadable with NVIDIA NeMo.",
+    )
+    parser.add_argument(
+        '--dataset_path', type=str, default='hf-audio/open-asr-leaderboard', help='Dataset path. By default, it is `hf-audio/open-asr-leaderboard`'
+    )
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        required=True,
+        help="Dataset name. *E.g.* `'librispeech_asr` for the LibriSpeech ASR dataset, or `'common_voice'` for Common Voice. The full list of dataset names "
+        "can be found at `https://huggingface.co/datasets/hf-audio/open-asr-leaderboard`",
+    )
+    parser.add_argument(
+        "--split",
+        type=str,
+        default="test",
+        help="Split of the dataset. *E.g.* `'validation`' for the dev split, or `'test'` for the test split.",
+    )
+    parser.add_argument(
+        "--device",
+        type=int,
+        default=-1,
+        help="The device to run the pipeline on. -1 for CPU (default), 0 for the first GPU and so on.",
+    )
+    parser.add_argument(
+        "--batch_size", type=int, default=32, help="Number of samples to go through each streamed batch.",
+    )
+    parser.add_argument(
+        "--max_eval_samples",
+        type=int,
+        default=None,
+        help="Number of samples to be evaluated. Put a lower number e.g. 64 for testing this script.",
+    )
+    parser.add_argument(
+        "--streaming",
+        action="store_true",
+        help="Stream the dataset lazily over the network instead of downloading it in full before the evaluation. Off by default for reproducible benchmark timings.",
+    )
+    parser.add_argument(
+        "--data_cache_root",
+        type=str,
+        default=None,
+        help="Root directory for caching audio files. Defaults to 'audio_cache' in current directory.",
+    )
+    args = parser.parse_args()
+
+    main(args)