Pith. sign in

REVIEW 3 major objections 6 minor 15 references

Practical Speech Recognition with HTK

T0 review · 3 major / 6 minor · reviewed 2026-08-14 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read This paper reports a practical HTK workflow for small-vocabulary speech recognition and a 100% command completion rate on an Indonesian voice-controlled electrical switch.

desk verdict A practical HTK tutorial with real field experience, but its 100% accuracy claim is circular because Section IV.G retrains on test errors. read the letter →

arxiv 1908.02119 v1 pith:M3ODAPTP submitted 2019-08-06 eess.AS cs.HCcs.LGcs.SD

Zulkarnaen Hatala This is my paper · ORCID
classification eess.AScs.HCcs.LGcs.SD
keywords speechrecognitionHTKsmall-vocabularyASRhomeautomationIndonesianlanguagenoiserobustnessMFCChiddenMarkovmodels
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper lays out a practical, step-by-step recipe for building a small-vocabulary speech recognizer with the Hidden Markov Model Toolkit (HTK) and applies it to an Indonesian voice-controlled electrical switch. The central idea is to make training and testing share the same acoustic world: training audio is recorded through the same HVite decoder, the same Bluetooth microphone, and the same room noise that will be present during live use. Two practical fixes carry much of the weight: scaling down microphone volume so ordinary noises do not trigger HTK's silence detector, and adding labeled noise patterns to the training database as dictionary words. The paper reports a 100% command completion rate for the resulting switch-control system. If the recipe works as described, it offers a fast, low-cost route to accurate command recognition for small vocabularies in moderately noisy homes.

What carries the argument

The load-bearing mechanism is a matched-condition training loop built around HTK's HVite program: HVite records the training audio, HCopy extracts MFCC_0_D_A features, HTK's HInit and HRest train phone-level HMM-GMM acoustic models, and HVite decodes live commands. Because the same decoder, microphone, room, and gain setting appear in training and testing, the model does not have to generalize across acoustic conditions. Supporting mechanisms are the volume-scaling rule $\hat{x}(t)=c x(t)$ with $0<c<1$, labeled NOISE entries in the pronunciation dictionary, and contraction entries such as NUM_3_4 to prevent word deletion.

What would settle it

Record a fresh set of Indonesian switch commands from the same speaker in the same room with the same microphone, run the finished HTK system on them without adding any of them to the training database or dictionary, and count completions. A completion rate below 100% would show that the reported figure reflects memorized training examples rather than generalization to new utterances.

Watch

Extended reading notes

Core claim

The core claim is that HTK, used with the right data-collection discipline, is enough to build a noise-tolerant small-vocabulary recognizer quickly. The decisive move is to record the training database with the same program that will perform live testing, HVite, under the same microphone, room, and gain conditions, instead of relying on clean speech corpora recorded in quiet studios. The paper further shows that reducing microphone gain by a factor $c$ with $0<c<1$ prevents everyday noises from being misread as words, and that including noise segments in the training database as explicitly labeled entries such as NOISE makes the recognizer robust to them. Indonesian cross-word contractions are handled by adding combined dictionary entries such as NUM_3_4. With this workflow, the paper finds that the Indonesian switch recognizer completed 100% of spoken commands, whether spoken as a switch number or as a named switch.

Load-bearing premise

The 100% completion rate assumes the test commands were not also training examples; the paper's refinement step, however, adds each failed sentence to the training database and retrains until it is recognized.

Editorial extensions

If this is right

  • A small-vocabulary recognizer for Indonesian can be built and deployed on an ordinary PC or smartphone with low-cost tools and a cheap Bluetooth microphone.
  • Recording training audio through the same decoder used for testing, in the same room, is the main route to reliable command recognition under noise.
  • Lowering microphone gain is enough to stop everyday noises from triggering HTK's silence detector and creating spurious words.
  • Adding dictionary entries for cross-word contractions removes a common word-deletion error in Indonesian digit sequences.
  • Residual online errors can be corrected by recording the failed sentence, labeling it, retraining, and adding variant pronunciations to the dictionary.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A fair multi-speaker evaluation would probably lower the completion rate, because the procedure as written ties the model to one speaker, one microphone, and one room; extending it would require per-speaker retraining or speaker adaptation.
  • The dictionary-entry fix for contractions is a workaround that grows combinatorially with vocabulary size, so a larger Indonesian recognizer would need a real lexicon or language model rather than more dictionary entries.
  • Because the underlying HTK algorithms are standard, the recipe's real contribution is data-collection discipline; a direct test would compare this matched-condition loop against a clean-trained baseline on the same held-out commands to quantify the gain.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper proposes a practical procedure for building a small-vocabulary automatic speech recognition (SVASR) system with the Hidden Markov Model Toolkit (HTK). The procedure covers grammar and pronunciation dictionary design, recording a training database under conditions matched to the target environment, MFCC feature extraction with optional PCA, HMM-GMM acoustic model training, online testing with HVite, a refinement step that retrains on newly observed errors, and deployment via an Arduino-based electrical switch. The method is illustrated with an Indonesian-language voice-controlled electrical switch (IDSwitch) system. The central empirical claim, stated in the abstract and conclusion, is a 100% command completion rate for this system.

Significance. If the central claim were validated, the paper would offer a useful practical recipe for rapidly building small-vocabulary ASR systems in moderately noisy environments. The paper gives explicit credit for several practical ideas: representing noise as a word in the dictionary, scaling microphone volume to reduce silence-detector false triggers, modeling cross-word contractions via dictionary entries, and recording training data with the same program (HVite) used for testing. These are concrete and potentially helpful for practitioners. However, the reported 100% command completion rate is not supported by a valid evaluation. The refinement procedure described in Section IV.G explicitly adds each failed test utterance to the training database, retrains the model, and retests the same utterance until it is recognized; the paper also permits adding the wrongly detected phoneme sequence to the dictionary. Consequently, the reported accuracy measures memorization of the test items rather than generalization to unseen commands.

major comments (3)
  1. [Section IV.G] The central claim of 100% command completion rate is invalid as an accuracy measurement because the evaluation is circular. Section IV.G states that when an error occurs in online testing, the sentence is taken, labelled, added to the database, and then retested after retraining. This means the utterances used to measure the final accuracy were included in the training set of the model that produced the reported result. The same section also permits adding wrongly detected phoneme sequences to the pronunciation dictionary, so the lexicon is adapted to the test items as well. The reported 100% therefore reflects the system's ability to memorize the specific test sentences, not its ability to generalize to unseen commands. To support the claim, the authors must freeze the acoustic model and dictionary after training and then evaluate on a separate set of previously unseen recordings, with no refinement applied after the evaluation begins.
  2. [Section IV.F and IV.H] The evaluation is under-specified in load-bearing ways. No number of test utterances, number of speakers, number of repetitions per command, or per-condition accuracy breakdown is given. The deployment section mentions a 'frame probability' threshold that revalidates recognition results, but the threshold value is never reported, so the reader cannot know how many ambiguous results were discarded before computing the 100% completion rate. Without these details, even a held-out evaluation would not be interpretable. The authors should report the test protocol, the test set size, the threshold used, and accuracy per command and per noise condition.
  3. [Section IV.D] The feature extraction parameters as written are internally inconsistent with the stated 16 kHz sampling rate. The text says 'windowing of 250 ms (400 samples)' and 'striding of 100ms (160 samples) and overlapping of 150 ms (240 samples)'. At 16 kHz, 400 samples correspond to 25 ms, not 250 ms, and 160 samples correspond to 10 ms, not 100 ms. This appears to be a factor-of-ten error that affects the reproducibility of the described procedure. The authors should correct the stated window length, frame shift, and overlap values to be consistent with the sampling rate.
minor comments (6)
  1. [Section III.F vs IV.G] The refinement step is described twice (Section III.F 'Refinement' and Section IV.G 'Refining') with slightly different wording; the relationship between these descriptions should be clarified to avoid confusion about when refinement is applied.
  2. [Table II] Table II has duplicate row numbers ('5' appears twice) and skips numbers 3 and 4. The table should be renumbered consecutively.
  3. [Figure 3] The caption contains a typo: 'Pronounciation' should be 'Pronunciation'.
  4. [Figure 5 caption] The caption text 'contains only one single sensor' presumably means 'sensor' rather than 'censor'.
  5. [Author biography] The biography says 'he thought computer science' but should read 'he taught computer science'.
  6. [References] Several references are incomplete or lack page numbers (e.g., [11] and [14]); the authors should verify that all citations follow a consistent format.

Circularity Check

1 steps flagged · score 8.0 of 10

The reported 100% command completion rate is produced by retraining on each failed test utterance until it is recognized, so the accuracy claim measures memorization rather than generalization.

  1. fitted input called prediction [Section IV.G Refining; echoed in Section VI Conclusion]
    "If in online testing an error occurs in the sentence that is spoken, then the sentence is taken and labelled and added to the database, then retest the sentence again. After that, it is examined whether the previous sentence can be recognized correctly now. The results showed that this method succeeded in correcting the errors of online testing."

    The paper's central claim is the '100% command completion rate' reported in the abstract and conclusion. That rate is achieved only after a refinement loop in which every utterance that fails online testing is labeled, inserted into the training database, and the model is retrained; the same utterance is then retested until it is recognized. The test utterances are therefore part of the training data by the paper's own description, so the final accuracy figure is a measure of the model's ability to reproduce previously seen examples, not its ability to generalize to unseen commands. The same section also allows adding the wrongly detected phoneme sequence to the dictionary, adapting the lexicon to the very test errors.

full rationale

The paper is mostly a practical tutorial on building an HTK-based Indonesian small-vocabulary ASR, and most of its methodological statements are not circular: grammar design, MFCC feature extraction, HMM-GMM training, and environment-matched recording are standard practices described without importing the result. The circularity is concentrated in the empirical claim. The abstract and conclusion assert a '100% command completion rate' for the IDSwitch system, but Section IV.G explicitly describes retraining on every sentence that failed online testing and retesting that same sentence until it is recognized. Under that protocol, the final test set is included in the training set, so the reported 100% rate is a fitting outcome of the refinement loop rather than an independent measure of recognition performance. The paper also states that unsuccessful entries can be added to the pronunciation dictionary as explicit words, which adapts the recognizer's lexicon to the test errors. No held-out partition, cross-validation, or evaluation on previously unseen recordings is reported, and no data or code are provided for external verification. Because the central accuracy result reduces by construction to the refinement procedure described in the paper's own text, a high circularity score is appropriate; this is not merely a self-citation issue or a difference from standard consensus.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The paper contributes a workflow, not a derivation. The central accuracy claim depends on several unreported tuning parameters and on an evaluation protocol that trains on the test set through the refinement loop; none of these are independently validated.

free parameters (6)
  • Volume scaling factor c = not reported (stated 0<c<1)
    Equation in Fig. 6 scales microphone signal by c; value is tuned to reduce noise triggering and never quantified in the text.
  • MFCC window/hop sizes = 250 ms window, 100 ms stride, 150 ms overlap (400/160/240 samples at 16 kHz)
    Chosen by hand in Section IV.D; no optimization or comparison.
  • PCA dimension after MFCC_0_D_A = not specified (only 'reduced')
    Section III.C and IV.D claim PCA improves accuracy and reduces computation, but the retained dimension is not reported.
  • Frame probability threshold for command acceptance = not specified
    Section IV.H says an upper layer uses a threshold on HVite frame probability to discard low-confidence commands; the threshold is never given.
  • HMM state and Gaussian mixture counts = not specified
    Section IV.E trains HMM-GMM with HInit/HRest but no topology (number of states, mixture components) is reported.
  • Number of training utterances per word/noise label = not specified
    Recording procedure in Section IV.C gives no counts per word or noise pattern, so the training set composition is unknown.
assumptions (5)
  • domain assumption HMM-GMM acoustic models are an adequate representation for sub-word speech recognition.
    Section II.B and IV.E adopt HTK's HMM-GMM framework without testing alternatives.
  • domain assumption MFCC features (with delta and acceleration) capture speech information needed for this task.
    Section III.C uses MFCC_0_D_A; the paper relies on the standard feature extraction literature.
  • domain assumption Matched training and testing environments and microphone are necessary and sufficient for good performance.
    Section III.B and V state this as the 'main key' of the procedure; it is a domain assumption about acoustic conditions.
  • ad hoc to paper Noise can be represented as a word with a phoneme sequence in the dictionary.
    Section IV.B defines NOISE with phone sequences like 't sp'; this is a modelling choice specific to this paper and not justified by experiments.
  • ad hoc to paper Test utterances used to measure the 100% rate are independent of the training data.
    This is the load-bearing evaluation assumption; Section IV.G undermines it by adding online errors to the training database and retraining until they pass.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Practical Speech Recognition with HTK." pith.science (2026). https://pith.science/paper/M3ODAPTP

@misc{pith2026190802119,
  author       = {Pith},
  title        = {Pith review of: Practical Speech Recognition with HTK},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/M3ODAPTP}},
  note         = {Machine review of arXiv:1908.02119}
}
read the original abstract

The practical aspects of developing an Automatic Speech Recognition System (ASR) with HTK are reviewed. Steps are explained concerning hardware, software, libraries, applications and computer programs used. The common procedure to rapidly apply speech recognition system is summarized. The procedure is illustrated, to implement a speech based electrical switch in home automation for the Indonesian language. The main key of the procedure is to match the environment for training and testing using the training data recorded from the testing program, HVite. Often the silence detector of HTK is wrongly triggered by noises because the microphone is too sensitive. This problem is mitigated by simply scaling down the volume. In this sub-word phone-based speech recognition, noise is included in the training database and labelled particularly. Illustration of the procedure is applied to a home automation application. Electrical switches are controlled by Indonesian speech recognizer. The results show 100% command completion rate.

Figures

Figures reproduced from arXiv: 1908.02119 by the authors.

Figure 1
Figure 1. Research method for SVASR IV. ILLUSTRATION The following example illustrates the development of SVASR with the proposed procedure. The application that is exemplified is building a recognition system for the Indonesian voice based electrical switcher (IDSwitch) and planting it into home automation electrical circuit switches. A. Prerequisites softwares Besides preparing computer and Bluetooth microphones, supporting… view at source ↗
Figure 4
Figure 4. Contraction modeling in dictionary Contraction modelling with dictionary entry is effective enough for SVASR, but could be a problem when applied to LVASR because of huge entries in order to accommodate all possibilities of contractions. C. Speech database a) Voice recording: sound is recorded using HSLab and HVite software with bluetooth microphone input devices. Other recording softwares such as Audacity and CoolE… view at source ↗
Figure 2
Figure 2. Grammar for IDSwitch And for the dictionary of the spelling of the sub-word pronunciation dictionary following the phoneme list in [6] is quoted in part as in Fig.3 [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figures from the paper (5 more)
Figure 3
Figure 3. Figure 3: Pronounciation dictionary untuk IDSwitch So on the dictionary at Fig.3 above the word is placed in the first column and the pronunciation is placed in the next columns. Sub-word sp at the end of some entries in the dictionary above stands for short pause which is if si…
Figure 9
Figure 9. Figure 9: HTK programs for creating HMM acoustic model F. Testing Tests were carried out with the HVite program. Hvite can perform testing offline using a recorded database or online directly from the microphone. In this case Bluetooth microphone is used. G. Refining Repair is d…
Figure 8
Figure 8. Figure 8: Noise marking D. Feature Extraction Feature extraction to convert waveform to MFCC (Mel Frequency Cepstral Coefficient) format. In this case the raw waveform sound is recorded from the replay buffer of HVite or using HSLab program recording. Feature extraction is done …
Figure 11
Figure 11. Figure 11: Embedded board with Atmega 328p and ESP 8266 The implementation diagram of the whole system is presented in figure 12 [PITH_FULL_IMAGE:figures/full_fig_p005_11.png]
Figure 12
Figure 12. Figure 12: Voice based electrical switcher implementation diagram V. DISCUSSION This procedure using HTK, has a dependency by requiring the same environmental conditions for testing and training. It also requires similarity of speakers and of microphone input devices. The desire…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

15 extracted references · 13 canonical work pages

  1. [1]

    The HTK Book version 3.5 alpha,

    S. Young, E. Gunnar, G. Mark, T. Hain, and D. K ershaw, “The HTK Book version 3.5 alpha,” Cambridge University, 2015

    2015

  2. [2]

    Phone Recognition on the TIMIT Database,

    C. Lopes and F. Perdigão, “Phone Recognition on the TIMIT Database,” 2009

    work page 2009

  3. [3]

    A large vocabulary continuous speech recognition system for Indonesian language,

    D. P. Lestari, K. Iwano, and S. Furui, “A large vocabulary continuous speech recognition system for Indonesian language,” 15th Indones. Sci. Conf. Japan Proc. , pp. 17–22, 2006

    work page 2006

  4. [4]

    Robust Sp eech Recognition for Car Environment Noise,

    H. Kokubo, A. Amano, and N. Hataoka, “Robust Sp eech Recognition for Car Environment Noise,” Electron. Commun. Japan, Part 3 , vol. 85, no. 11, pp. 2190–2197, 2002

    work page 2002

  5. [5]

    Illustration of the IPA: Indonesian,

    C. D. Soderberg and K. S. Olson, “Illustration of the IPA: Indonesian,” J. Int. Phon. Assoc. , vol. 38, no. 2, pp. 209–213, 2008

    work page 2008

  6. [6]

    Speaker-Independent Phon e Recognition Using Hidden Markov Models,

    K. Lee and H.-W. Hon, “Speaker-Independent Phon e Recognition Using Hidden Markov Models,” IEEE Trans. Acoust. , vol. 37, no. 11, pp. 1641–1648, 1989

    work page 1989

  7. [7]

    The SPAR Speech Fil ing System,

    M. A. Huckvale, D. M. Brookes, L. T. Dworkin, M . E. Johnson, D. J. Pearce, and L. Whitaker, “The SPAR Speech Fil ing System,” Eur. Conf. Speech Technol. , pp. 305–308, 1987

    work page 1987

  8. [8]

    Speak and unSpea k with Praat,

    P. Boersma and V. van Heuven, “Speak and unSpea k with Praat,” Glot Int. , vol. 5, no. 9–10, pp. 341–347, 2001

    work page 2001

Show all 15 references
  1. [9]

    The CMU ARCTIC Speech Databases,

    K. John and A. W. Black, “The CMU ARCTIC Speech Databases,” in 5th ICSA Speech Synthesis Workshop - Pittsburg , 2004, pp. 223–224

    2004

  2. [10]

    Comparison of Parametric Representations for Monosyllabic Word Recognition i n Continuously Spoken Sentences,

    S. B. Davis and P. Mermelstein, “Comparison of Parametric Representations for Monosyllabic Word Recognition i n Continuously Spoken Sentences,” IEEE Trans. Acoust. , vol. 28, no. 4, pp. 357–366, 1980

    1980

  3. [11]

    A tutorial on Principal Componen ts Analysis,

    L. I. Smith, “A tutorial on Principal Componen ts Analysis,” 2002

    2002

  4. [12]

    PCA-Based Speech E nhancement for Distorted Speech Recognition,

    T. Takiguchi and Y. Ariki, “PCA-Based Speech E nhancement for Distorted Speech Recognition,” J. Multimed. , vol. 2, no. 5, pp. 13–18, 2007

    2007

  5. [13]

    A Tutorial on Hidden Markov Mo dels and Selected Applications in Speech Recognition,

    L. R. Rabiner, “A Tutorial on Hidden Markov Mo dels and Selected Applications in Speech Recognition,” in Proceedings of the IEEE , 1989, vol. 77, no. 2, pp. 257–286

    1989

  6. [14]

    ATmega328P 8-bit AVR Microcontroll er with 32K Bytes In-System Programmable Flash DATASHEET

    Microchip, “ATmega328P 8-bit AVR Microcontroll er with 32K Bytes In-System Programmable Flash DATASHEET.”

  7. [15]

    ESP8266EX Version 6.0 Data sheets,

    Espressif Systems, “ESP8266EX Version 6.0 Data sheets,” 2018. Zulkarnaen Hatala was born in Ambon, on 19 Agustus 1977. He received Sarjana Teknik on Informatics at 2002 and Master Teknik on Telecommunications at 2005. Both degress are from Telkom University, Bandung Indonesia....

    2018

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.