AmericasNLP 2026 · 2026 · Workshop

Anita Baral et al. (Miami University) · → Paper · Demo: ? · Code: ?

Builds the first neural TTS for Myaamia (Miami-Illinois), a dormant-then-reclaimed Algonquian language, by training FastSpeech, Glow-TTS and VITS from scratch on 8.18 hours of community-archived recordings, with VITS giving the lowest spectral and pitch error and expert listeners rating its naturalness 3.05 against 4.79 for the original recordings.

Problem

Myaamia lost its last first-language speakers in the mid-20th century and is now being reclaimed by the Miami Tribe of Oklahoma and the Myaamia Center at Miami University. Learners depend on archived audio, and an estimated 95% of the available recordings come from just two people, which limits access to spoken models of the language. No neural TTS system existed for Myaamia, and the more recent zero-shot, diffusion and flow-matching systems the authors survey need far more data and compute than the community archive provides and are not well supported in open-source toolkits for low-resource languages. The paper also sets out to do this work under Indigenous data sovereignty, with the Myaamia Center retaining ownership and governance of all data.

Method

Data. The corpus (named Myaamia-TTS in the paper) is 14,358 dictionary recordings (about 10.4 hours) made at the Myaamia Center since 2010, drawn from the Indigenous Languages Digital Archive (ILDA). It covers three categories: basic word forms, command (imperative) forms and example sentences. Seven speakers appear, but two contribute roughly 95% of the audio. Utterances are short: 59.6% have 3 to 14 characters and the mean duration is 2.61 seconds (Table 2). The authors converted stereo MP3s to mono, removed silences longer than 0.5 s below -40 dB, normalised volume, and resampled everything to 22,050 Hz (the sources mixed 44.1, 22.05 and 48 kHz). They then split the corpus 80/20 at random by utterance, giving 11,486 training utterances (8 h 18 min) and 2,872 test utterances (2 h 4 min). The split is not stratified by speaker.

Input representation. Myaamia uses a phonemic orthography with 20 basic symbols (four vowels, each short or long, plus 12 consonants), so the models take characters directly as input with no grapheme-to-phoneme step. Long vowels are written as doubled letters. Myaamia Center linguists authored and verified the transcripts, and the models learn alignment internally from utterance-level pairs.

Models. Three existing architectures from the Coqui TTS toolkit are trained from scratch, all with Coqui’s default hyperparameters (no hyperparameter search), batch size 32, for 1,000 epochs on a single NVIDIA A30:

  • FastSpeech: a non-autoregressive feed-forward Transformer encoder and decoder, trained with Adam (lr 1e-4) and a NoamLR schedule.
  • Glow-TTS: a normalizing-flow acoustic model with a relative-positional Transformer encoder, trained with RAdam (lr 1e-3) and 4,000 NoamLR warmup steps.
  • VITS: an end-to-end conditional VAE with adversarial (GAN) training. It has a six-layer text encoder (two attention heads, 768 FFN channels) and is trained with AdamW (lr 1e-3) in fp16. Mel loss gets weight 45.0, alongside the KL, generator, discriminator and duration losses.

The paper does not say how the seven speakers are handled (for example, whether speaker IDs are used), and it does not report parameter counts. Training took 27.69 h for FastSpeech, 33.87 h for Glow-TTS and 60.10 h for VITS (§3.3).

Evaluation. Objective metrics are computed on all 2,872 test utterances after DTW alignment to the reference recording: MCD (excluding the 0th coefficient), F0 RMSE and duration RMSE. Only VITS goes through the subjective evaluation. Four Myaamia experts took part: two are the main speakers in the training data and two are Myaamia linguists. Each rated 20 samples on 5-point scales, half of them synthesized and half original recordings. Ten samples were rated for naturalness, intelligibility and overall quality, and ten for intonation/rhythm and linguistic accuracy. The experts’ written feedback feeds a qualitative error analysis (§4.2, §4.3).

Key Results

Objective comparison (Table 3, all lower is better).

MetricFastSpeechGlow-TTSVITS
MCD19.82 ± 2.3317.91 ± 2.1915.22 ± 2.66
F0 RMSE30.51 ± 29.9320.00 ± 23.7317.50 ± 20.69
Duration RMSE0.31 ± 0.370.30 ± 0.360.29 ± 0.40

VITS has the lowest mean on all three metrics, and the ordering VITS < Glow-TTS < FastSpeech holds for MCD and F0 RMSE. The duration RMSE means differ by only 0.01 to 0.02, well inside the standard deviations, so the three systems are effectively tied on timing. The authors note that all MCD and F0 RMSE values are higher than those usually reported for well-resourced TTS, though they do not compare against a specific reference figure. No significance tests are reported.

Inference cost (§3.3, CPU, all 2,872 test utterances). FastSpeech is the fastest with an RTF of 7.43 ± 3.00, Glow-TTS reaches 8.59 ± 3.29, and VITS is the slowest at 12.71 ± 4.53. All three run slower than real time on CPU.

Expert ratings of VITS (Table 4, MOS ± SD).

DimensionVITSOriginal recordings
Naturalness3.05 ± 0.894.79 ± 0.42
Understandability3.05 ± 1.104.55 ± 0.69
Overall quality3.20 ± 0.774.75 ± 0.44
Intonation/rhythm3.52 ± 0.814.52 ± 0.56
Linguistic quality3.40 ± 0.914.45 ± 0.56

The largest shortfall against the original recordings is in naturalness (VITS 1.74 points lower). The smallest is in intonation/rhythm (1.00 lower) and linguistic quality (1.05 lower).

Cross-language comparison (Table 5). The paper puts its naturalness MOS (3.05, 8.18 h of training data) next to published figures for Võro TTS (3.62, 17 h) and Mundari TTS (3.69, 24.76 h) and calls its result competitive given less data. These numbers come from different languages, listener pools and test protocols, so the comparison is only indicative. Myaamia TTS has the lowest naturalness score of the three.

Error analysis (§4.3). The experts flagged several errors:

  • Phonemic contrasts collapsed: š was realised as the affricate c (/tš/), the cluster /mhs/ drifted toward /nk/, and /l/ was heard as /w/.
  • Consonant clarity was generally reduced.
  • Contrastive vowel length was wrong in several words.
  • Long multi-word utterances had glitches, skips and choppy syllable boundaries.

None of these issues were reported for the original recordings.

Novelty Assessment

The contribution is the application of existing tools rather than a new method. Three standard Coqui TTS recipes are trained with default settings on a new, non-public Indigenous-language corpus. What is new for the field is the language itself, the documented preprocessing of a heterogeneous archival dictionary corpus, a first objective benchmark for Myaamia, and the expert error analysis. That analysis ties specific synthesis failures to features of the language, namely phonemic vowel length written as doubled characters and preaspirated or clustered consonants. The data-governance framing (CARE principles, IRB protocol, all data kept on institution-controlled compute) is a substantive part of the paper but not a technical novelty. The corpus is not released, so the benchmark cannot be reproduced outside the Myaamia Center.

Field Significance

low — The paper adds a single-language case study to the evidence that small-footprint VITS training works on a few hours of archival Indigenous-language data. It also gives a concrete list of character-input failure modes (vowel-length and consonant-contrast errors) that a phoneme-aware front end might address. The human evaluation does not raise the level. It covers only VITS, uses four raters (two of whom are the training-data speakers), reports no inter-rater agreement and compares to other languages only across studies, so the three-architecture comparison rests on objective metrics from one non-public test split.

Claims

  • supports: In a low-resource single-language setting with a few hours of archival speech, an end-to-end VAE-GAN TTS model reproduces spectral and pitch characteristics more faithfully than two-stage non-autoregressive or flow-based acoustic models trained on the same data.

    Evidence: Trained on 8.18 h of Myaamia with Coqui defaults, VITS has MCD 15.22 and F0 RMSE 17.50, against 17.91 and 20.00 for Glow-TTS and 19.82 and 30.51 for FastSpeech. Duration RMSE is effectively tied (0.29 to 0.31). (§4.1, Table 3)

  • complicates: The quality advantage of end-to-end VAE-GAN TTS comes with higher training and CPU inference cost than simpler feed-forward or flow-based acoustic models, which matters for deployment in community education settings.

    Evidence: VITS needed 60.10 h of training and runs at CPU RTF 12.71, against 27.69 h and RTF 7.43 for FastSpeech and 33.87 h and RTF 8.59 for Glow-TTS. None of the three runs in real time on CPU. (§3.3)

  • complicates: Character-level input works as a baseline for a language with a phonemic orthography, but it can still lose phonemic contrasts and contrastive vowel length that the orthography encodes with digraphs or doubled letters.

    Evidence: Experts heard š realised as /tš/, /mhs/ reduced toward /nk/, /l/ perceived as /w/, and wrong vowel lengths. The authors link the vowel errors to long vowels being written as doubled characters and point to the lack of phoneme-based input. (§4.3, §5)

  • complicates: In very small language communities, subjective TTS evaluation often has to rely on a handful of expert raters who may include the training-data speakers, which limits how far perceptual results generalise and makes cross-study MOS comparisons weak.

    Evidence: Only VITS was rated, by four experts (two of them the main recording speakers), with no inter-rater agreement reported. The naturalness MOS of 3.05 is set against Võro (3.62) and Mundari (3.69) scores from separate studies. (§4.2, §5, Tables 4-5)

Limitations and Open Questions

Narrow perceptual evidence

Only VITS was rated by listeners, using four experts (two of whom recorded most of the training data), 20 samples per rater and no inter-rater agreement. The ranking of the three architectures therefore rests only on DTW-aligned objective metrics from one non-public, randomly split test set.

  • The random utterance-level split is not stratified by speaker or lexical item. Because the data is dictionary recordings dominated by two speakers, near-duplicate forms built on the same stem may appear in both training and test sets. The paper does not examine this.
  • Because two speakers contribute about 95% of the data, the synthesized voice mostly reflects those two people, and the handling of the other five speakers is not described.
  • All three systems use Coqui default hyperparameters with no tuning, so the architecture comparison may partly reflect how well each default recipe suits short utterances.
  • There is no explicit prosody, stress or pitch modelling, and the authors report artifacts in longer multi-word utterances. Since 95% of training utterances have 25 or fewer characters, it is unclear whether the system can handle sentence-length material.
  • No inference ran in real time on CPU, which matters for the planned integration into the Šaapohkaayoni learning portal. Compression or distillation is proposed but not tried.
  • Recent pretrained or cross-lingual systems were not tested, so it is unknown whether transfer from a related or high-resource language would close the gap to the original recordings.

Wiki Connections

  • Transformer Encoder-Decoder TTS — Trains FastSpeech (a feed-forward Transformer encoder and decoder) from scratch on Myaamia, where it trails both Glow-TTS and VITS on MCD and F0 RMSE.
  • GAN Vocoder — Trains VITS, whose adversarially trained end-to-end waveform generation gives the lowest spectral and pitch error of the three systems on 8.18 h of data.
  • Subjective Evaluation — Runs a small expert MOS test of VITS against original recordings on five dimensions, and documents the rater-pool constraints of very small language communities.
  • Shipibo-Konibo TTS — A previous AmericasNLP low-resource Indigenous TTS case study that the paper cites as part of the neural TTS work it builds on.
  • Multilingual TTS for Ojibwe, Mi’kmaq, and Maliseet — A related Algonquian-language TTS system cited as prior work. This paper instead trains monolingual Myaamia models.
  • FastSpeech 2 — Cited as the successor to the FastSpeech baseline. The paper trains the original FastSpeech, not FastSpeech 2.
  • F5-TTS — Cited as an example of the flow-matching systems that the authors consider too data- and compute-hungry for this setting, and not evaluated.
  • VALL-E — Cited as an example of the neural codec language models that the authors set aside because of their data requirements.
  • NaturalSpeech 2 — Cited as an example of the diffusion-based TTS that the authors set aside for this low-resource setting.