Unexpected attention results using phoneme based cleaners

[lexkoro]

Branch: Master (540d811dd5)

Hi,

trying to train a multi-speaker model using the current master branch with phoneme computation I encounter the following weird attention alignment problem.

I am using the basic_german_cleaner for text cleaning. I also tried the https://github.com/repodiac/german_transliterate from @repodiac but it yielded the same results. Corresponding config config.zip

Switching back to character based training resolves the problem. Any ideas?

[george-roussos]

I just had a look at the german cleaner and it looks okay -- what dataset are you using? If you try e.g. echo "lächeln" | phonemize -l de on the terminal, do you get the correct output?

I would also try a run with restoring the German single speaker TTS. It should help with alignments

[thorstenMueller]

We used german cleaner by @repodiac during training „thorsten“ dataset and didn’t encounter these weird alignment issues. Even running tts-server with german cleaner shows no issues on pronouncing german umlauts.

I’ve just generated a test sentence and uploaded it on soundcloud.

https://soundcloud.com/thorsten-mueller-395984278/tts-thorsten-de-testsatz-mit-deutschen-umlauten

Dieser Testsatz dient zur Überprüfung der Aussprache der deutschen Umlaute mit deutschem Bereinigungsskript, wie Ä in Änderungswunsch, Ü in Überraschungsei, Ö in Öffentlichkeit oder auch Straße mit ß sollte gehen. 

[repodiac]

Hi @SanjaESC, if you can nail it down to a wrong phoneme encoding based on cleaning I am happy to help (and fix) out :) As @thorstenMueller pointed out, we do not know those issues yet - if you find a wrong transliteration I can also only strongly encourage you to file an issue on my github page: https://github.com/repodiac/german_transliterate/issues

[lexkoro]

@george-roussos Using echo "Die süße Hündin läuft in die Höhle des Bären, der sie zum Kaffeekränzchen eingeladen hat." | phonemize -l de yields the following result diː zyːsə hʏndɪn lɔøft ɪn diː høːlə dɛs bɛːrən dɛɾ ziː tsʊm kafeːkɾɛntsçən aɪnɡəlɑːdən hat which seems fine to me. I also printed out the phoneme conversion while it was caching the initial run and they all seemed fine.

@thorstenMueller @repodiac Pretty sure your cleaner isn't the root cause of the problem, since I get the same results with the default cleaner. I'll try some more experiments and report back.

PS: The character based multi-speaker model is training just fine.

[thorstenMueller]

@SanjaESC my output of phonemize -l de is near, but little bit different than your output:

thorsten: diː zyːsə hʏndɪn lɔøft iːn diː høːlə deːs bɛːrən deːɾ ziː tsʊm kafeːkɾɛntsçən aɪnɡəlɑːdən hɑːt

sanja: diː zyːsə hʏndɪn lɔøft ɪn diː høːlə dɛs bɛːrən dɛɾ ziː tsʊm kafeːkɾɛntsçən aɪnɡəlɑːdən hat

[lexkoro]

@thorstenMueller Interesting! Any idea why that could be? Did you make any changes to espeak? My phonemizer version is 2.2.1.

Update: The attention aligns using a single speaker.

[erogol]

you might be using GMM attention. Can you post your config.json?

[lexkoro]

you might be using GMM attention. Can you post your config.json?

// AUDIO PARAMETERS
"audio":{
    // stft parameters
    "fft_size": 1024,         // number of stft frequency levels. Size of the linear spectogram frame.
    "win_length": 1024,      // stft window length in ms.
    "hop_length": 256,       // stft window hop-lengh in ms.
    "frame_length_ms": null, // stft window length in ms.If null, 'win_length' is used.
    "frame_shift_ms": null,  // stft window hop-lengh in ms. If null, 'hop_length' is used.

    // Audio processing parameters
    "sample_rate": 22050,   // DATASET-RELATED: wav sample-rate.
    "preemphasis": 0.0,     // pre-emphasis to reduce spec noise and make it more structured. If 0.0, no -pre-emphasis.
    "ref_level_db": 20,     // reference level db, theoretically 20db is the sound of air.

    // Silence trimming
    "do_trim_silence": true,// enable trimming of slience of audio as you load it. LJspeech (true), TWEB (false), Nancy (true)
    "trim_db": 60,          // threshold for timming silence. Set this according to your dataset.

    // Griffin-Lim
    "power": 1.5,           // value to sharpen wav signals after GL algorithm.
    "griffin_lim_iters": 60,// #griffin-lim iterations. 30-60 is a good range. Larger the value, slower the generation.

    // MelSpectrogram parameters
    "num_mels": 80,         // size of the mel spec frame.
    "mel_fmin": 0.0,        // minimum freq level for mel-spec. ~50 for male and ~95 for female voices. Tune for dataset!!
    "mel_fmax": 8000.0,     // maximum freq level for mel-spec. Tune for dataset!!
    "spec_gain": 20.0,

    // Normalization parameters
    "signal_norm": true,    // normalize spec values. Mean-Var normalization if 'stats_path' is defined otherwise range normalization defined by the other params.
    "min_level_db": -100,   // lower bound for normalization
    "symmetric_norm": true, // move normalization to range [-1, 1]
    "max_norm": 4.0,        // scale normalization to range [-max_norm, max_norm] or [0, max_norm]
    "clip_norm": true,      // clip normalized values into the range.
    "stats_path": null    // DO NOT USE WITH MULTI_SPEAKER MODEL. scaler stats file computed by 'compute_statistics.py'. If it is defined, mean-std based notmalization is used and other normalization params are ignored
},

// VOCABULARY PARAMETERS
// if custom character set is not defined,
// default set in symbols.py is used
"characters":{
    "pad": "_",
    "eos": "~",
    "bos": "^",
    "characters": "ABCDEFGHIJKLMNOPQRSTUVWXYZÄÖÜabcdefghijklmnopqrstuvwxyzäöüß!'(),-.:;? ",
    "punctuations":"!'(),-.:;? ",
    "phonemes":"iyɨʉɯuɪʏʊeøɘəɵɤoɛœɜɞʌɔæɐaɶɑɒᵻʘɓǀɗǃʄǂɠǁʛpbtdʈɖcɟkɡqɢʔɴŋɲɳnɱmʙrʀⱱɾɽɸβfvθðszʃʒʂʐçʝxɣχʁħʕhɦɬɮʋɹɻjɰlɭʎʟˈˌːˑʍwɥʜʢʡɕʑɺɧɚ˞ɫ"
},

// DISTRIBUTED TRAINING
"distributed":{
    "backend": "nccl",
    "url": "tcp:\/\/localhost:54321"
},

"reinit_layers": [],    // give a list of layer names to restore from the given checkpoint. If not defined, it reloads all heuristically matching layers.

// TRAINING
"batch_size": 32,       // Batch size for training. Lower values than 32 might cause hard to learn attention. It is overwritten by 'gradual_training'.
"eval_batch_size":16,
"r": 7,                 // Number of decoder frames to predict per iteration. Set the initial values if gradual training is enabled.
"gradual_training": [[0, 7, 64], [1, 5, 64], [50000, 3, 32], [130000, 2, 32], [290000, 2, 32]], //set gradual training steps [first_step, r, batch_size]. If it is null, gradual training is disabled. For Tacotron, you might need to reduce the 'batch_size' as you proceeed.
"loss_masking": true,         // enable / disable loss masking against the sequence padding.
"ga_alpha": 10.0,        // weight for guided attention loss. If > 0, guided attention is enabled.
"apex_amp_level": null, // level of optimization with NVIDIA's apex feature for automatic mixed FP16/FP32 precision (AMP), NOTE: currently only O1 is supported, and use "O1" to activate.

// VALIDATION
"run_eval": true,
"test_delay_epochs": 400,  //Until attention is aligned, testing only wastes computation time.
"test_sentences_file": "/home/akorolev/master/projects/TTS/sentences.txt",  // set a file to load sentences to be used for testing. If it is null then we use default english sentences.

// OPTIMIZER
"noam_schedule": false,        // use noam warmup and lr schedule.
"grad_clip": 1.0,              // upper limit for gradients for clipping.
"epochs": 1000,                // total number of epochs to train.
"lr": 0.0001,                  // Initial learning rate. If Noam decay is active, maximum learning rate.
"wd": 0.000001,                // Weight decay weight.
"warmup_steps": 4000,          // Noam decay steps to increase the learning rate from 0 to "lr"
"seq_len_norm": false,         // Normalize eash sample loss with its length to alleviate imbalanced datasets. Use it if your dataset is small or has skewed distribution of sequence lengths.

// TACOTRON PRENET
"memory_size": -1,             // ONLY TACOTRON - size of the memory queue used fro storing last decoder predictions for auto-regression. If < 0, memory queue is disabled and decoder only uses the last prediction frame.
"prenet_type": "bn",           // "original" or "bn".
"prenet_dropout": false,       // enable/disable dropout at prenet.

// TACOTRON ATTENTION
"attention_type": "original",  // 'original' or 'graves'
"attention_heads": 4,          // number of attention heads (only for 'graves')
"attention_norm": "sigmoid",   // softmax or sigmoid.
"windowing": false,            // Enables attention windowing. Used only in eval mode.
"use_forward_attn": false,     // if it uses forward attention. In general, it aligns faster.
"forward_attn_mask": false,    // Additional masking forcing monotonicity only in eval mode.
"transition_agent": false,     // enable/disable transition agent of forward attention.
"location_attn": true,         // enable_disable location sensitive attention. It is enabled for TACOTRON by default.
"bidirectional_decoder": false,  // use https://arxiv.org/abs/1907.09006. Use it, if attention does not work well with your dataset.
"double_decoder_consistency": true,  // use DDC explained here https://erogol.com/solving-attention-problems-of-tts-models-with-double-decoder-consistency-draft/
"ddc_r": 7,                           // reduction rate for coarse decoder.

// STOPNET
"stopnet": true,               // Train stopnet predicting the end of synthesis.
"separate_stopnet": true,      // Train stopnet seperately if 'stopnet==true'. It prevents stopnet loss to influence the rest of the model. It causes a better model, but it trains SLOWER.

// TENSORBOARD and LOGGING
"print_step": 25,       // Number of steps to log training on console.
"tb_plot_step": 100,    // Number of steps to plot TB training figures.
"print_eval": false,     // If True, it prints intermediate loss values in evalulation.
"save_step": 10000,      // Number of training steps expected to save traninpg stats and checkpoints.
"checkpoint": true,     // If true, it saves checkpoints per "save_step"
"tb_model_param_stats": false,     // true, plots param stats per layer on tensorboard. Might be memory consuming, but good for debugging.

// DATA LOADING
"text_cleaner": "basic_german_cleaners",
"enable_eos_bos_chars": false, // enable/disable beginning of sentence and end of sentence chars.
"num_loader_workers": 2,        // number of training data loader processes. Don't set it too big. 4-8 are good values.
"num_val_loader_workers": 2,    // number of evaluation data loader processes.
"batch_group_size": 0,  //Number of batches to shuffle after bucketing.
"min_seq_len": 6,       // DATASET-RELATED: minimum text length to use in training
"max_seq_len": 150,     // DATASET-RELATED: maximum text length

// PATHS
"output_path": "/home/akorolev/master/projects/TTS/Trainings/",

// PHONEMES
"phoneme_cache_path": "/home/akorolev/master/projects/TTS/phoneme_cache/",  // phoneme computation is slow, therefore, it caches results in the given folder.
"use_phonemes": true,           // use phonemes instead of raw characters. It is suggested for better pronounciation.
"phoneme_language": "de",     // depending on your target language, pick one from  https://github.com/bootphon/phonemizer#languages

// MULTI-SPEAKER and GST
"use_speaker_embedding": true,      // use speaker embedding to enable multi-speaker learning.
"use_external_speaker_embedding_file": false, // if true, forces the model to use external embedding per sample instead of nn.embeddings, that is, it supports external embeddings such as those used at: https://arxiv.org/abs /1806.04558
"external_speaker_embedding_file": null, // if not null and use_external_speaker_embedding_file is true, it is used to load a specific embedding file and thus uses these embeddings instead of nn.embeddings, that is, it supports external embeddings such as those used at: https://arxiv.org/abs /1806.04558
"use_gst": true,                    // use global style tokens
"gst":  {                           // gst parameter if gst is enabled
    "gst_style_input": null,        // Condition the style input either on a
                                    // -> wave file [path to wave] or
                                    // -> dictionary using the style tokens {'token1': 'value', 'token2': 'value'} example {"0": 0.15, "1": 0.15, "5": -0.15}
                                    // with the dictionary being len(dict) <= len(gst_style_tokens).
    "gst_embedding_dim": 512,
    "gst_num_heads": 4,
    "gst_style_tokens": 10
},

// DATASETS
"datasets":   // List of datasets. They all merged and they get different speaker_ids.
    [
        {
            "name": "gothic_tts",
            "path":  "/home/akorolev/master/projects/data/SpeechData/tts/DialogSpeech/",
            "meta_file_train": null,
            "meta_file_val": null
        }
    ]

[erogol]

Try changing prenet_type:original. And switch it to bn after some training.

[lexkoro]

Try changing prenet_type:original. And switch it to bn after some training.

Sorry maybe I should have stated that initially I tried training it with prenet_type: original and had the same results. Using a single-speaker + phonemes works on the other hand.

What confuses me is that training on graphemes works just fine. There the attention aligns already at around ~3k steps.

I will test some more and report back if I find something. Could be obv. also a problem on my end. ^^

[george-roussos]

Try a run restoring Thorsten's TTS. It may help 👍

[lexkoro]

Update:

Using the current dev branch works. Still not sure what the root of the caused problem is.

Not sure what causes the problem. Char-based training works fine. Single-Speaker training with phonemes also works fine. Only Multi-speaker doesn't want to align.

Will close this for now.