inesc-id/camoes_asr

CAMÕES: ASR Benchmark for European Portuguese

CAMÕES is a comprehensive and diverse evaluation benchmark for European Portuguese (EP) Automatic Speech Recognition (ASR): 46.2 h of test data, 14 corpora, 1,311 speakers aged 3 to 100, covering five domains from read to highly spontaneous speech.

📄 Paper: CAMÕES: A Comprehensive Automatic Speech Recognition Benchmark for European Portuguese (ASRU 2025) 🤗 Models and resources: ASR for European Portuguese collection 📝 Submit your model: submission portal

This repository does not contain audio data. It is the central hub for the benchmark: description, evaluation protocol and leaderboard.

Domains

From lower to higher spontaneity:

• Read Speech (RS): read audiobooks and text prompts, such as news articles, speech commands, numbers, single words and digits, with little to no spontaneity.
• Broadcast News (BN): news content from public Portuguese TV channels: mostly read or planned speech with a specific type of enunciation, uttered by professionals.
• Talks/Lectures (T/L): TEDx talks and university lectures: prepared but not read, with a higher degree of spontaneity than the previous domains.
• Conversational Speech (CS): celebrity interviews and other recordings from Portuguese TV channels.
• Sociolinguistic Interviews (SI): highly spontaneous conversational speech recorded in various Portuguese regions and social contexts, often with poor recording conditions and strong accents.

Leaderboard

WER (%) on the CAMÕES benchmark. EP Avg. is the mean of the five domain averages (each domain has equal weight); the ranking uses EP Avg. AAP = African and Asian Portuguese, BP = Brazilian Portuguese (Table II of the paper). Training data: FT = fine-tuned on EP-425 (425 h of EP); TFS = trained from scratch on EP-425; PT-All = EP-425 + BP-417 + AAP-8; AAP = EP-425 then AAP-8; BP = BP-417 only; 0-shot = used as released. Bold = best, underline = second best (EP columns).

Rank	Model	# Trainable params	RS	BN	T/L	CS	SI	EP Avg.	AAP	BP
1	WhisperLv3-X (FT)	1.55B	7.4	4.7	11.3	11.2	27.9	12.5	24.0	27.2
2	WhisperLv3-X (PT-All)	1.55B	7.9	4.7	12.3	11.6	28.6	13.0	23.3	18.8
3	WhisperLv3-X (AAP)	1.55B	8.2	5.9	12.1	11.7	28.9	13.4	22.7	26.3
4	EBranch-w2vBERT2 + 4-gram LM (TFS)	114M	8.0	5.4	15.6	13.4	27.1	13.9	26.6	41.9
5	EBranch-w2vBERT2 (AAP)	114M	8.3	5.5	16.0	13.7	27.4	14.2	26.3	41.9
6	EBranch-w2vBERT2 (TFS)	114M	8.3	5.4	16.0	14.9	27.2	14.4	26.7	42.4
6	EBranch-w2vBERT2 (PT-All)	114M	8.7	5.5	16.4	13.5	28.0	14.4	24.6	20.7
8	EBranch-XLSR (TFS)	114M	9.6	6.5	16.7	18.2	29.3	16.1	27.6	48.1
9	WhisperLv3 (FT)	1.55B	7.2	4.6	13.6	14.9	43.2	16.7	101.4	28.2
10	EBranch (TFS)	114M	9.4	6.5	18.0	16.5	35.4	17.2	36.3	59.0
11	WhisperLv3-X (0-shot)	-	16.4	8.2	16.6	15.3	39.3	19.2	29.0	24.6
12	Phi-4-MI (0-shot)	-	15.5	8.6	17.9	21.9	44.5	21.7	27.1	25.9
13	Canary-1b-v2 (0-shot)†	-	24.0	10.9	16.8	19.5	45.1	23.3	-	-
14	Phi-4-MI (FT)	1.3B	9.6	7.2	16.7	24.4	59.5	23.5	35.6	31.1
15	WhisperLv3-X (BP)	1.55B	17.2	13.8	24.1	20.8	46.6	24.5	29.4	18.8
16	WhisperLv3 (0-shot)	-	32.4	7.9	15.4	18.3	49.0	24.6	29.9	25.8
17	Nemotron-3.5-ASR-streaming-0.6b, 1.12 s latency (0-shot)‡	-	27.5	17.6	23.9	24.5	54.1	29.5	-	-
18	OWSM-CTC v4 (0-shot)	-	22.5	24.4	32.0	28.7	52.1	31.9	37.1	32.3
19	SeamlessM4T-v2 (0-shot)	-	26.7	17.3	26.3	27.9	64.5	32.5	46.4	33.3
20	Nemotron-3.5-ASR-streaming-0.6b, 160 ms latency (0-shot)‡	-	32.4	25.2	29.3	29.9	63.5	36.1	-	-
21	MMS-all (0-shot)	-	33.9	25.7	40.3	38.2	65.4	40.7	57.5	50.5
22	EBranch-w2vBERT2 (BP)	114M	37.9	32.6	42.2	40.0	54.9	41.5	38.3	21.3

All results except † and ‡ are from Table III of the paper. † Evaluated after publication, EP only. ‡ Evaluated after publication by the CAMÕES authors, EP only: NVIDIA's streaming model decoded with NeMo's cache-aware streaming inference (target_lang=pt-PT, greedy decoding) at two latency settings. WhisperLv3-X = Whisper large-v3 decoded with WhisperX (voice-activity detection before transcription). Linked models are released on the Hub.

Evaluation protocol

• Metric: word error rate (WER, %), computed at corpus level per test set.
• Text normalisation: reference and hypothesis are normalised with the CAMÕES Portuguese normaliser, pt_normalizer.py, before scoring (lower-case, punctuation removed, numbers spelled out in European Portuguese, disfluency and annotation markers removed).
• Averaging: each domain score is the mean of its corpora; EP Avg. is the mean of the five domain scores, so every domain counts equally regardless of its size.
• Orthography: reference transcriptions follow the spelling from before the 1990 Orthographic Agreement (pré-Acordo), e.g. projecto, director.

Scoring one test set:

from jiwer import wer
from pt_normalizer import EuPtTextNormalizer  # pt_normalizer.py from this repository

normalizer = EuPtTextNormalizer()
pairs = [(normalizer(ref), normalizer(hyp)) for ref, hyp in zip(references, hypotheses) if normalizer(ref)]
print(f"WER: {100 * wer([r for r, _ in pairs], [h for _, h in pairs]):.2f}%")

How to participate

Submit your model through the official submission portal.

Model requirements: the model must be uploaded to the Hugging Face Hub and work with the pipeline("automatic-speech-recognition") function:

from transformers import pipeline

# Load your model from the Hugging Face Hub
asr_pipeline = pipeline("automatic-speech-recognition", model="your-org/your-model")

# Run inference
result = asr_pipeline("path/to/audio.wav")

Benchmark details

Domain	Corpus	Test (h)	Test (#spks)	Age	M/F (%)	Notes
RS	BD-Publico	2.0	10	18-28	50/50	Read sentences extracted from an EP newspaper.
	CommonVoice	1.8	42	13-59	48/12	pt-PT subset. Speaker count estimated from the client ids provided in the corpus.
	MLS_extended	1.0	10	NI	27/73	EP extension of MLS: automatically aligned audiobooks.
	PT_Adults	1.6	17	25-59	52/48	Read prompts. Corresponds to YMA.
	PT_Children	2.1	52	3-10	56/44	Corpus of child speech.
	PT_Elderly	1.3	172	60-100	26/74	Read prompts. Test speakers are aged 76-100.
	SpeechDat	9.7	604	14-98	46/54	Telephone speech sampled at 8 kHz, upsampled to 16 kHz.
BN	Alert	6.6	175	NI	70/29	Broadcast news.
T/L	Lectra	2.6	7	NI	NI	University lectures. Speakers are shared among partitions.
	MuAViC	0.4	2	NI	60/40	TEDx talks.
CS	Postport	3.9	>30	NI	54/24	Debates and entertainment (a few documentaries and information).
	VoxCelebPT	2.9	13	NI	38/62	Voices of Portuguese celebrities collected from YouTube.
SI	Fala Bracarense	6.1	8	15-92	45/55	Sociolinguistic interviews recorded in Braga, 2009-2014.
	PT Fundamental	4.2	169	17-69	44/56	Low-quality recordings of interviews collected in the 1970s.
Total		46.2	1,311

NI = information not available. M/F may not sum to 100% because of speakers with unknown gender. For more details on the corpora, the training data and the baselines, see the paper.

Citation

@inproceedings{camoes,
  title     = {{CAMÕES: A Comprehensive Automatic Speech Recognition Benchmark for European Portuguese}},
  author    = {Carvalho, Carlos and Teixeira, Francisco and Botelho, Catarina and Pompili, Anna and
               Solera-Ure{\~n}a, Rub{\'e}n and Paulo, S{\'e}rgio and Juli{\~a}o, Mariana and Rolland, Thomas and
               Mendon{\c{c}}a, John and Pereira, Diogo and Trancoso, Isabel and Abad, Alberto},
  booktitle = {Proceedings of the IEEE Automatic Speech Recognition and Understanding Workshop (ASRU)},
  year      = {2025}
}

Credits

• Curated by: Human Language Technologies Lab (HLT), INESC-ID Lisboa
• Funded by: Portuguese national funds through Fundação para a Ciência e a Tecnologia, under project UIDB/50021/2020 (DOI:10.54499/UIDB/50021/2020), by the Portuguese Recovery and Resilience Plan and NextGenerationEU European Union funds under project C644865762-00000008 (Accelerat.AI).
• License: Other (the test corpora keep their original licences)