Audio data collection for long-tail languages remains essential for building reliable automatic speech recognition (ASR) systems, despite recent advances in multilingual AI. While fine-tuning existing models can reduce data requirements, performance for low-resource languages still falls short of commercial expectations without high-quality transcribed audio. The bottom line: thoughtful, ethical audio data collection is not only a technical requirement but a foundation for reaching the other half of the world’s population who speak long-tail languages.
Recent research reinforces this point. Languages that have an associated language model, even a relatively simple n-gram model, achieve better ASR performance than those without one (Findings of ACL 2024). A language model allows the ASR system to check whether a transcription is plausible for the target language, increasing the likelihood that the output makes sense.
Interestingly, this language model need not be large; it is sufficient to contain a superset of the vocabulary present in the audio training data. This means that for current long-tail languages where text data may be limited, collecting additional transcribed audio is a more effective investment than building increasingly complex language models.
Your flywheel is only as fast as your annotation pipeline.
LXT embeds into your retraining loop, delivering labeled data at the cadence your model demands, not months later.
This context makes recent advances in multilingual ASR particularly compelling. In late 2025, Meta released Omnilingual ASR, providing speech recognition for approximately 1,200 languages at below 10% character error rate (CER). This represents a substantial portion of today’s approximately 7,000 extant languages (Ethnologue). But does this mean that ASR is a solved problem?
‘Solved’ Is a Loaded Term
Thinking about ‘solved’ problems, such as putting humans on the moon, creating breathable and waterproof fabric, preventing outbreaks of measles and smallpox, training random humans into chess masters, and many other examples, all lead to the conclusion that ‘solved’ means different things to different people.
On one level, the term ‘solved’ implies that the problem is no longer interesting because someone, somewhere, knows how to fix the problem. The term is also often used to suggest that a commercially available product already performs at an effective and reliable level. I’ll leave it as an exercise for the reader to determine whether they consider the above-mentioned problems to be ‘solved’, and in which way.
In the case of ASR for long-tail languages, the first meaningful steps toward normalising spoken interaction with computers have been taken, but human-level robustness and naturalness are still far from routine.
Fine-Tuning Long-Tail Languages
The definition of ‘solved’ becomes particularly clear when looking at fine-tuning ASR for long-tail languages. While adapting an existing model can significantly reduce data requirements, performance may still fall short of commercial expectations.
For example, fine-tuning Whisper first to Tamil and then to Malasar, a Dravidian language spoken by approximately 10,000 people in South India, resulted in word error rates exceeding 40%, even after punctuation normalisation (November 2024 paper on multistage fine-tuning for low-resource ASR). By contrast, models such as NVIDIA’s canary-qwen-2.5b, which was optimised for English (GitHub repository), achieve average WERs around 5.63%. This gap in accuracy highlights the limits of fine-tuning in the absence of sufficient, well-matched training data.
Cross-Linguistic Training with Omnilingual
One of Omnilingual’s most interesting contributions is its ability to train multiple languages concurrently. Meta updated wav2vec 2.0 in 2021 to learn discrete phonemes or even shorter segments, scaling it up to 7B parameters (Meta AI blog on Omnilingual ASR). Two decoder variants map these segments to characters. Crucially, the model was trained to identify cross-linguistic acoustic patterns that can be reused across languages, achieving state-of-the-art performance with substantially less data than earlier approaches.
Notably, related languages benefit most strongly from this setup. Although English and German sound quite distinct to human listeners, Omnilingual shows that they share enough acoustic structure to support mutual learning. This finding suggests that training multiple varieties of the same language, or closely related languages, may now be possible with similar or even lower total audio durations than were previously required for a single variety.
Demographics and Targeted Varieties
As companies expand ASR systems beyond a small set of high-resource languages, demographic considerations become increasingly important. Fine-tuning ASR models requires speakers from the target locales, and possibly includes both native and non-native speakers.
As ASR becomes reliable for ‘mainstream’ varieties of English, Spanish, French, and German, users of non-mainstream varieties (including Austrian German and Australian English) reasonably expect to be understood as well. Applying consistent data-sourcing practices and quality metrics across both high-resource and long-tail languages helps ensure that expansion does not come at the cost of performance or user trust.
The Politics of Audio Data
Large-scale ASR is inseparable from the politics of data. As noted in the Omnilingual paper, questions of how data is gathered, who is compensated, and who controls its use are central to the success of multilingual speech technology.
For many low-resource languages, the absence of trained developers who speak the language limits local ownership of AI tools. Thoughtful, ethical audio data collection is therefore not only a technical requirement, but also a social and economic one.
The Bottom Line in 2026
Fine-tuning existing models has changed the economics of ASR for long-tail languages, but it has not eliminated the need for high-quality audio data. Training or adapting models for multiple related languages is now more feasible than ever, yet achieving reliable, real-world performance still requires careful consideration of the end use case and multiple iterations of data collection and evaluation.
Progress has changed the rules, but audio data remains the foundation.
